Sie sind auf Seite 1von 8

A Review of Wind Power and Wind Speed Forecasting Methods With Different Time Horizons

Saurabh S. Soman, Hamidreza Zareipour, Senior Member, IEEE, Om Malik, Life Fellow, IEEE, and Paras Mandal, Member, IEEE

Abstract In recent years, environmental considerations have prompted the use of wind power as a renewable energy resource. However, the biggest challenge in integrating wind power into the electric grid is its intermittency. One approach to deal with wind intermittency is forecasting future values of wind power production. Thus, several wind power or wind speed forecasting methods have been reported in the literature over the past few years. This paper provides insight on the foremost forecasting techniques, associated with wind power and speed, based on numeric weather prediction (NWP), statistical approaches, artificial neural network (ANN) and hybrid techniques over different time-scales. An overview of comparative analysis of various available forecasting techniques is discussed as well. In addition, this paper further gives emphasis on the major challenges and problems associated with wind power prediction. Index Terms Artificial neural network, auto-regressive moving average, fuzzy logic, numeric weather prediction, persistence method, statistical methods, wind speed forecasting, wind power forecasting.

I. INTRODUCTION ntermittency of wind is the biggest challenge to implementing wind-energy a reliable autonomous source of electric power. Large-scale wind-penetration requires answers to a lot of problems such as competitive market designs, realtime grid operations, standards of interconnection, ancillary service requirements and costs, quality of power, capacity of transmission system and its future upgrades, stability and reliability of power system, optimal reductions in green house gas emissions of entire power system typically determined by optimal amount of wind penetration into system [1], [2]. Improved wind-forecasting is known as an efficient tool to overcome many of these problems. For example, when it comes to competitive electricity markets, accurate wind forecast is always alluring for a variety of reasons. Firstly, appropriate incentives of attractive market price are offered on energy imbalance charges based on market price. Secondly, a correct forecast can help to develop well-functioning hourahead or day-ahead markets [1], [3]. Some of the mentioned issues are further addressed in [4][7]. Improving wind power prediction has significant economic and technical advantages. A probabilistic method for calculating energy expenses associated with prediction

errors for wind generators is proposed in [4] where case studies show that error prediction costs can reach as much as 10% of the total incomes from selling wind energy. A shortterm probabilistic forecast of wind power is presented in [5] where a general method for optimal bidding strategy based on uncertainty information of forecasts is discussed. Fossil fuel savings of around 15% to 25% that could be achieved as compared to persistence model by a large scale integration of varying amounts of wind power using the numeric weather predictor (NWP) model in conjunction with a physical flow model and a statistical model as forecasting method is illustrated in [6]. The possibility of using a wind farms output for ancillary services, like reactive power support, primary frequency support, power oscillation damping support, etc., in case when wind farms output remains unutilized due to forecast errors is reported in [7] and a cost model is developed for that. The finding from this model is that the higher wind forecast errors may lead to increased payment to wind farms for their reactive power service; due to the increased lost opportunity cost. This paper provides a detailed review on wind power and wind speed forecasting based on recent available published papers. The major contribution of this paper is the classification of the literature on wind power/speed forecasting according to their forecasting horizons and time-scales. While there are several review papers available on wind power/speed forecasting (e.g., [8], [9]), to the best of the authors, a review with this specific focus does not yet exist in the literature. The significance of this review is that it provides a clear comparison and review of each methodology based on its intended time horizon. The rest of this paper is organized as follows: Section II discusses the classification of various forecasting techniques based on time-scales followed by an overview of basic forecasting methods in section III. The difference between wind-speed and wind-power forecast is discussed in section IV. The several of forecasting approaches presented in section V is evaluated in section VI and points out the best available based on their apparent performance and results. Finally, conclusions are drawn in section VII.

TABLE I TIME-SCALE CLASSIFICATION FOR DIFFERENT FORECASTING TECHNIQUES Time horizon Very shortterm Range Few seconds to 30 minutes ahead 30 minutes to 6 hours ahead 6 hours to 1 day ahead Applications - Electricity Market Clearing - Regulation Actions

TABLE II BASIC WIND SPEED AND POWER FORECASTING METHODS Forecasting Method Persistence Method/ Nave Predictor Physical Approach Subclass Examples Remarks - Benchmark approach - Very accurate for veryshort and short term. - Use of meteorological data such as wind speed and direction, pressure, temperature, humidity, terrain structure etc. - Accurate for long term. - Accurate for short-term - Their hybrid structures useful for medium to long term forecasts - Usually, outperform timeseries models - Accurate for short-term - Some very good timeseries models supersede NN structures. - Spatial correlation is good for short term. - Entropy based training of model improves performance. - Considering nonGaussian error pdf improves accuracy - ANFIS is very good for very-short term forecast. -NWP + NN structures are very accurate for medium and long-term forecasts.

P(t+k)=P(t) - Global Forecasting System - MM5 - Prediktor - HIRLAM, etc. - Feed-foreward - Recurrent - Multilayer Perceptron - Radial Basis Function - ADALINE, etc. - ARX - ARMA - ARIMA - Grey Predictors - Linear Predictions - Exponential Smoothing, etc. - Spatial Correlation - Fuzzy Logic - Wavelet Transform - Ensemble Predictions - Entropy based training, etc. - NWP+NN - ANN + Fuzzy logic = ANFIS - Spatial Correlation + NN - NWP+timeseries

Short-term

- Economic Load Dispatch Planning - Load Increment/Decrement Decisions - Generator Online/Offline Decisions - Operational Security in Day-Ahead Electricity Market - Unit Commitment Decisions - Reserve Requirement Decisions - Maintenance Scheduling to Obtain Optimal Operating Cost

Numeric Weather Predictors (NWP)

Mediumterm

Long-term

1 day to 1 week or more ahead

Artificial Neural Network (ANN) Statistical Approaches Time-series Models

Note: Forecasting term limits are not strictly defined and some relaxation may be granted depending on the application of the forecasting model.

II. TIME-SCALE CLASSIFICATION Time-scale classification of wind forecasting methods is vague. However, as shown in Table I, wind forecasting can be separated into three categories: Very short-term forecasting: From few seconds to 30 minutes ahead. Short-term forecasting: From 30 minutes to 6 hours ahead. Medium-term forecasting: From 6 hours to 1 day ahead. Long-term forecasting: From 1 day to 1 week ahead. Table I also presents the applications of specific time horizon in view of the operation of electricity systems. III. OVERVIEW OF WIND POWER FORECASTING METHODS A general overview of wind forecasting methods is presented in Table II. Most common wind forecasting techniques developed and reported in literature use one of the followings: 1) Persistence Method This method is also known as Nave Predictor. It is assumed that the wind speed at time t+t will be the same as it was at time t. Unbelievably, it is more accurate than most of the physical and statistical methods for very-short to shortterm forecasts. Industry still uses it for very-short term forecasts [1], [10]. Hence, any forecast method that is developed should, first, be tested against classical benchmark of persistence method to check how much it can improve over this technique [11]. 2) Physical Approach Physical systems use parameterizations based on a detailed physical description of the atmosphere. Usually, wind speed given by the weather service on a coarse grid is transformed to the onsite conditions at the location of the wind farm [12]. Numeric Weather Prediction (NWP): This method is a physical approach to wind forecasting. NWP models operate

Some New Techniques

Hybrid Structures

by solving complex mathematical models that use weather data like temperature, pressure, surface roughness and obstacles. NWPs are rendered on supercomputers as they need lots of computations. Customarily, NWPs are run 1 or 2 times a day due to the difficulty to gain information in short-time and the associated high costs. This limits its usefulness to medium to long-term forecasts (> 6 h ahead). These give most accurate predictions when weather conditions are stable [10], [13]. 3) Statistical Approach The statistical approach is based on training with measurement data and uses difference between the predicted and the actual wind speeds in immediate past to tune model parameters [10], [12]. It is easy to model, inexpensive, and provides timely predictions. It is not based on any predefined mathematical model and rather it is based on patterns. Errors are minimized if patterns are met with historical ones. Subclassification of this approach is: Time-series based models, and neural network (NN) based methods. Auto-Regressive Moving Average (ARMA) models are the most popular type in the time-series based approach to predict future values of wind speed or power. Several variations are auto regressive integrated moving average (ARIMA), seasonal- and fractional-ARIMA, ARMA with exogenous

input (ARMAX or ARX). Few other time-series models are grey predictors, linear predictions, exponential smoothing, etc. The NNs are trained using past data taken over a long timeframe to learn the relationship between input data and output wind-speeds. NNs have an input layer where historical data is fed for learning, hidden layer(s) and an output layer providing forecast results. NN models can be feed-forward neural networks (FNNs), multi-layer perceptrons (MLP), recurrent neural networks (RNNs), radial basis function (RBf) NNs, Adaline networks, etc. Generally, NNs outperform time-series models for almost all time-scales, although this is not necessarily universal. For example, s-ARIMA and Adaline NN models are used to forecast wind speed in Mexico and are compared with each other [14]. Results show s-ARIMA better follows the actual pattern. Similarly in [15], single exponential smoothing is used for forecasting, that follows actual trend very well with values of error adjustment coefficient 0.9 . Comparison with the same Adaline NN shows that exponential smoothing model presents better sensitivity to the adjustment & prediction of the wind speed. However, in both, when the number of training vectors is increased for the given NN model, its performance is improved. 4) Hybrid Approach In general, combination of different approaches such as mixing physical and statistical approaches or combining shortterm and medium-term models, etc., is referred to as a hybrid approach. For example, radiative transfer and NN techniques are combined with Special Sensor Microwave/Imager (SSM/I) to get ocean surface wind speeds and direction in [16]. Results show that combination of NN would substantially enhance the impact of these data in NWPs than just using SSM/I. One of the newly popular techniques is the model based on the spatial correlation of wind speeds. Here, spatial correlation between the wind speeds at different sites is taken into account. The wind speed time-series of local point and its neighboring sites are employed to predict the wind speed usually by NNs or adaptive neuro-fuzzy networks [17], [18]. It is because the change in wind speed time-series observed at remote station seen at measuring station with some time delay. IV. WIND SPEED VERSUS WIND POWER Power output of a wind turbine depends on the wind speed, which varies with time and depends on regional weather patterns and type of landscape. Relationship between wind speed (m/s) through swept area A (m2) of wind turbine and wind energy per unit time or wind power P (W) is 1 3 P = A , where is density of air (kg/m3), which rests 2 on temperature and pressure of air. From this relationship, it can be seen that the relationship between wind speed and power is nonlinear, basically cubic. Any error in wind speed forecast will actually give a large (cubic) error in wind power. In addition, for entire wind farm, this relation is more complex as different turbines in the farm use multiple wind directions and speed to achieve optimal power output of wind farm.

Thus, small error in wind speed forecast can generate larger error in wind power forecast. To map wind speed into power, easy way is to use manufacturers power curve for each wind turbine separately and aggregate the results; as in [19]-[22]. But it doesnt give optimal result. Therefore, the best approach as described in [1] is to use a power curve created using measured wind at the site. This can ameliorate Root Mean Squared Error (RMSE) of forecast by 20% compared to only using manufacturers power curve. V. REVIEW OF WIND POWER FORECASTING TECHNIQUES This section is divided in four parts based on the timescales, and for each of them, various forecasting techniques used in recent years are discussed briefly as below. A. Very-Short Term Forecasting A few papers are available for very-short term forecasting timeframe. A case study from Tasmania, Australia for veryshort term (2.5 minutes ahead) forecasting using Adaptive Neuro-Fuzzy Interface System (ANFIS) to forecast wind vectors rather than wind speed or power is described in [10] where it is reported that wind direction is more important than wind speed over very-short term interval to gain most from the forecast. To train the model, data-set providing a 21 month time series in steps of 2.5 minutes is used as input and speed and direction are converted into two vectors u and v. Results show that ANFIS produced less than 4% mean absolute percentage error (MAPE) whereas that for persistence is approximately 30% in deciding either u or v. It is also suggested to develop very-short term forecast models for slightly longer time duration, because many deregulated markets are cleared every 5 minutes and settlements are done every 30 minutes. This is reported in [3] for the timeframe of 5 to 15 minutes. Tests are conducted with actual wind speed data of 30s resolution from a wind farm to check the portability of system. It is suggested, for future extension, to develop a system that forecasts for 30 min timeframe, and to consider the conditions like non-uniform wind speeds in a single park. A method based on linear prediction model of ARMA in conjunction with filtering of wind waveforms for prediction of wind speed is described in [23]. Output is the linear combination of present and past samples. As lesser frequency components in the spectrum of a signal provide better shortterm prediction of signal, wind speed is filtered by a low pass filter imposed by wind turbine mechanical system. Prediction results of 1s and 5s ahead are reported in [25]. B. Short Term Forecasting Most research on wind forecasting has been done in this time scale. Short-term forecasting of 1-h ahead is investigated in [24] where author presents Bayesian approach to develop an AR model of 6th order. Comparison with persistence model indicates usefulness of method but lower order AR models fail to provide sufficient accuracy. ARMA with only historical generation data for 6-h advance

predictions is examined in [11]. The proposed methodology beats persistence by 7% in first hour and 18% in sixth hour. Results of test of model for 10-min ahead period produce unsatisfactory performance showing that the ability of ARMA models will differ when applied to different time periods. A hybrid model combining nine alternative statistical forecasting methods utilizing a case-study of a Mexican wind farm is presented in [25]. Final prediction is made by an aggregated single model using adaptive linear combination of alternative methods, where weight of each model is based on its actual forecast performance. Results show that final forecast of power is nearly as same as its real values. ARX and NN models against persistence method for the forecasting horizons of 1, 3, 6 and 12 hour are compared in [26]. Tests show persistence outperforming ARX for less than 13 hours; while NN model is better than persistence model. RNN for scheduling autonomous wind-diesel system for next 2 hours is used in [27]. Three different NN configurations are tested based on RMSE criterion. All improve over persistence and the NN model considering forecast errors of all time steps and giving multiple outputs for each ahead-time step(s) simultaneously is implemented in control system, which increased profits. A Mexican case study applying NN to the hourly time series prediction is described in [28]. A model for every month of year is developed. Four different NN configurations are tested. Simplest model with two layers, two input and one output neurons turned out to be the best with 0.0016 MSE and 0.0399 Mean Absolute Error (MAE) values. Examples of [27] & [28] show that the selection of a suitable NN model requires careful analysis and depends on the characteristic of problem. A Gaussian Process (GP) with kernel-machine technique and Bayesian estimation for estimating upper and lower bounds and average of wind speed of next 1-hour is proposed in [29]. Historical wind speed data with other meteorological variables is simulated and compared against RBf and MLP NN methods. Results show improvement of as much as 27% of average error and 13% of maximum error. Spatial correlations for 15-min to 3-hour ahead speed forecasting using NN for predicting relationship between data measured at 1 reference and 2 remote sites are used in [17]. Inputs are last few 1 to 5-min speed averages from remote sites and back propagation (BP) algorithm is used for training. Results, using cases of long and short spatial distances, show forecast efficiency improved by 28% against the persistence, implying that data from neighboring sites is always useful. Another application of spatial correlation method using TSK fuzzy interface model for 2-hour ahead prediction is presented in [18]. Inputs are wind speed at upwind station and direction and training was done using genetic algorithm. Case of flat terrain shows improvement of 29% based on Fitness function criteria over nave predictor, whereas in case of complex terrain application, the model fails due to non-correlation between local and remote sites. Three different NNs namely BP, adaptive linear element, and RBF for 1-hour ahead wind speed prediction are compared based on three different error criteria MAE, RMSE

and MAPE in [30]. Tests were performed with varying number of observations provided as input and with different learning rates. Results show, for each NN model, different best configurations were found for different sites and error criteria; i.e., none of NNs outperform others universally. This suggests that the type of NN model to be used, number of model inputs & learning rates should be carefully decided. A universal evaluation criteria & robust method are needed for combining forecasts from various models to avoid inconsistency. A new approach, known as Mycielski, is introduced in [31]. In this, next value in currently ongoing random process is the longest repeating data chain that had shown up in the past data sequence. Three Turkish case studies show maximum RMSE of only 1.5%. Due to such robustness, model can be adapted easily for forecasting in any region. In [32], data is pre-processed in order to obtain properties like standard deviation, average and slope before feeding it to either fuzzy logic or NN. This reduces fuzzy rule base or NN size and speeds its learning process. Results employing RMSE and coefficient of determination (COD) criteria show less computational time, less RMSE and more COD for the proposed methodology. A new method based on Grey model (GM) for 1-hour ahead forecasting is introduced in [33]. In this, the original data series is transformed into a new series of less noise and randomness. After that, a differential equation is formed and its coefficients are determined using least square method. Forecasted values of new series are calculated using these coefficients, and actual forecasted values of original series are retrieved by inverse operation. But it presents overshoots and so, to reduce them, alpha GM is presented in [34], that has weighting factors for calculation of coefficients of differential equation. But it has poor time-series tracking feature. So, two new models Improved GM and Averaged GM are developed. Improved GM generates two shifted-prediction models from the normal GM when it is at stage of forecasted new series and they are combined to generate final improved GM based on their weights. In averaged GM, these weights of each model are taken equal (that is, 0.5 each). Results based on MAE and RMSE show that all models improve over persistence model and average GM has overall superiority for forecasts amongst all GMs and nave predictor. A hybrid method is reported in [19] that uses wavelet method to decompose original time-series into a number of subseries and then, improved ARIMA method is used to predict the next values in each subseries. Results of subseries are aggregated and compared with classical time-series model and BP NN using MAE, Mean Square Error (MSE) and MAPE criteria. Suggested method gives better results (less error) compared to others for 3-step, 5-step and 10-step ahead prediction respectively. As expected, the bigger the forecasting steps are, the lower is the accuracy. An augmented complex statistics based linear sequential least mean square (ACLMS) algorithm is used in [35], for simultaneous modeling of wind speed and direction in the domain of complex numbers and forecasts wind power by

finite impulse response (FIR) filter. Results are checked against standard complex filter for finely sampled wind data over 1-, 3- and 6-hour intervals based on MAE and COD criteria, that show good improvement for ACLMS algorithm. C. Medium Term Forecasting Most of methods developed for this time-period are based on NN approaches, physical weather models, hybrid models combining both of these or some new techniques. Reference [36] uses wavelet transform to decompose the signal and cut up data into different frequency components. A prediction algorithm approximates coefficients of decomposed waveforms which are then reconstructed to contain original sample plus forecast of next 24 hours. Results based on MSE show that out of four different daubechies wavelets, only one or two outperformed persistence method. A case-study using FNN, trained by Levenberg-Marquardt (LM) method is presented in [37] for 1 to 24-hour ahead forecasts. Optimal number of neurons is found by trial & error method. Results give 12% less MAPE for FNN compared to persistence. Computation time taken was just 5 s. A hybrid approach for 1 to 48-hour ahead forecasts is developed in [20]. It provides preliminary forecast based on NWPs using RBf NN, estimates quality of forecast based on errors between predicted and actual values of power and direction, and poor NWPs are improved by fuzzy rules to obtain improved predictions. Results are compared based on MAE and RMSE against nave predictor and just NWP method, which show substantial improvement for both (up to 46% for persistence). Finally, authors also compare given model with the state-of-art statistical model, with NWPs and online-prediction data as input, proposed in [38]. Results show hybrid approach of [20] superseding state-of-art method. A new f-ARIMA model is examined in [39]. It is a special case of ARIMA process. Model parameters were estimated using exact maximum likelihood and optimized using Akaikes Information Criterion, which compares accuracy against complexity of model. Results based on COD, Daily Mean Error (DME), and variance showed very high COD (95%) and very low DME & variance compared to persistence and ARIMA. Translation from wind speed to wind power forecasts also gave accurate results for 24 to 48-hour ahead. Goal of [40] is to verify performance of a regional NWP, known as Eta model. Eta forecasts with the coarse and the nested, finer resolutions; where coarse forecasts are used as boundary conditions for the nested runs. Results of 12 to 36hour ahead wind speed forecasts are compared against wind observed at nearest surface station and at 10 m height at a site in Sweden, which show high COD (0.8), and low MAE and RMSE. Thus, Eta model is useful for wind energy modeling. A two-stage hybrid network is presented in [41]. A Bayesian clustering by dynamics classifies input training dataset into several subsets with similar properties in unsupervised manner. A support vector regression fits training data in each subset in supervised way. Past wind generation behavior is extrapolated by finding hidden recurring trends, structural changes etc. via RQA analysis for

1 to 48-hour ahead forecasts, given the meteorological information. It improves over persistent for all predictions by as much as 40% based on RMSE and MAE criteria. Usual assumption of Gaussian distribution of wind power forecast error is denied in [42]. Analysis of errors against persistence for 1 to 24-hour ahead forecast data shows that their kurtosis varies from 3 to 10, matching Beta probability distribution function (pdf); while that of Gaussian distribution is 3. To validate this, forecast results are sorted into several power bins and pdf within each bin is fit approximately on Beta pdf. Parameter pairs of fit Beta pdf of forecast error for every bin are obtained and final pdf is calculated by summing all pdfs by their weights, decided by probability of a forecast belonging to particular bin. Results for a 24-hour ahead forecast show Beta pdf fits actual data perfectly. A case-study for the sizing of Energy Storage System shows that given algorithm performs very well for values of loss more than 2%. D. Long Term Forecasting Long-term forecasting is very important in restructured electricity markets and its applications are mentioned earlier. A new type of physical method to predict pdf of wind power generation for 1-to-10 days ahead forecast using weather ensemble predictions (WEP) is developed in [43]. The model is calibrated and smoothened to properly estimate the uncertainty in weather conditions. Comparison of results with statistical time series methods (ARMA based models) show WEP gave more accurate results over a period of week. Three RNNs with numerical wind speed and direction forecasts as input from four nearby nodes are used in [44] for 72-hour ahead prediction. Results, compared with two FNNs and the nave predictor based on MAE and RMSE criteria, show RNNs improve over persistence by as much as 50% and follow real power curve more closely than FNNs. It still improves when additional node predictions are included. Method of [44] is very exhaustive and time-consuming. So, it is required for RNN to forecast task in lesser time and without requiring too much meteorological data or complications. This is done in reference [45] by applying RNN for wind speed forecasts of each month separately. RNN is trained by 1 year of data at only one site. Simulations show that RNNs are more accurate than FNNs for several hours to day-ahead forecasts. Fifth generation mesoscale (MM5) model using meteorological data from a global NWP as its initial and boundary conditions is presented in [46]. Data is interpolated to obtain mean hourly wind speed predictions at a certain wind farm in Spain. These along with wind direction and temperature are fed to a NN to get speed forecasts of 48-hour ahead for each wind turbine. Optimal number of neurons is found by trial and error. MAE and MSE errors are low (1.8 m.s-1 on average) and their frequency distribution lies within a narrow band near mean value. Graphs of actual and predicted speeds show that hybrid system follows actual series closely. In [21], two hybrid systems of MM5 and NNs are combined such that NNs act as transfer functions between NWP wind speed and direction forecast and corresponding electric power

generation; and last values related to measured wind power are used to obtain forecast in first hours. A system known as FORECAS uses single MLP NN; whereas other, SGP system, uses combination of Kalman filter, ARMA model and several NN models, all optimized using fuzzy system. Forecast results show that, both, FORECAS and SGP predict real mean power production quite accurately and very closely and improve by 48% on average over persistence as by RMS criterion for same real life case studies of various applications such as intraday market (12-24 hour ahead), day-ahead market (24-48 hour ahead) and maintenance scheduling (48-72 hour ahead). Number of FNN configurations such as 1 or 2 hidden layer, varying number of neurons (5 to 15) and different training algorithms like scaled conjugate gradient and LM for predicting mean monthly wind speed off 28 stations in Nigeria are tested in [47]. Input variables are latitude, longitude, altitude, and month of year. Optimal design of two hidden layers, 15 neurons each, and LM algorithm gave 8.9% MAPE, and 93.8% COD. Model is used to monitor and predict wind speed at locations with no monitoring stations, due to good results. For non-Gaussian error distribution, error criteria based on minimizing information content or entropy (as entropy is measure of information content) such as MEE (minimum error entropy), MCC (maximum correntropy) or MEEF (MEE with fiducial points) instead of minimizing its variance (MSE criterion) are more suitable for training of forecast models, as suggested in [48]. To verify this, tests based on MCC and MEEF criteria against MSE criterion were conducted on FNNs having one hidden layer of nine neurons for 72-hour ahead forecasts in half hour intervals. Inputs were wind speed and direction. For offline training, NNs trained with MEEF and MCC showed average 1.5% less difference of Normalized MAE over NNs trained with MSE. For online training, this was around 1.15%. Average difference between online and offline trained MEEF was 1.56%. Major improvements can be achieved if this approach is integrated into some more sophisticated wind power prediction systems. VI. DISCUSSION A. Design Issues Wind power forecasting is very specialized area. Usually, forecast model developed for one site doesnt suit the other. But generalization capability of forecast model can be improved by considering following issues: 1) Data Preprocessing: Generally, data to be used as input is fed as it is to the forecast model such as time-series, NN or fuzzy system. However, as shown in [32], just preprocessing the input data makes huge difference in performance of the model. Preprocessing means dividing input so that a local model may be designed for each subclass [49]. This may be done by extracting the properties like standard deviation, mean, variance, slope from the input data. This filters the irregularities in data and makes models simpler. Adopting this concept would help a lot.

2) Design and training of forecast models: Wind power forecasting models developed for one location usually dont match the other site due to variety of reasons like change in terrain, different wind speed patterns, different atmospheric factors such as temperature, pressure, humidity. So, developing a portable and easy-to-operate model is a big challenge. Such a method is presented in [41] that is strongly robust, easy to modify for various parks and so, quite practical and cheap. Further research on this track is encouraged. About training, excessive training should be avoided as otherwise models tend to over-parameterize or overfit the input data and this is true in case of both NNs and time-series models such as ARMA. It means model complexity and parameters are so large that it fits data too well and tries to track every single data point in the training set and so, ends up with poor forecasts in test data set. Thus, overfitting is result of excessive complexity and lack of clear guidelines on how many parameters in a model are too many, for a given sample size [49]. To avoid this problem, many researchers intuitively reduce the weight of number of neurons, either by omitting some of the neurons or by eliminating some connections. Among some other design issues are: a) Selection of number of input nodes This is crucial because unnecessary input data will rather slow down the process as it happened in [44] when including pressure and temperature. This may not always happen and many times inclusion of various weather variables is required for accurate forecasts. Also, if too old past values are used, they would be irrelevant to recent patterns, i.e., low auto-correlation will deteriorate forecast performance. Box method [49] or method (f-ARIMA) presented in [39] can be used for fitting timeseries models like ARIMA on suitable lags, selected analyzing auto-correlation function (ACF) or partial ACF. b) Selection of number of outputs It depends on whether a single forecast or multiple forecasts for a particular look-ahead time is needed. For multiple forecasts, multiple models or outputs can be used simultaneously but this arrangement increases the model complexity. Therefore, usually, single output is used and multiple forecasts are made one by one iteratively, as in most of the references here. 3) Implementation and Validation of forecast models: Implementation of a forecast model implies determining its parameters. Selecting optimal number of model parameters is a difficult task and techniques like trial and error method, Box method [49] or like those presented in [39] have been used widely in the papers under review for all type of models like ARMA, NN or fuzzy system. This is especially difficult for NNs or fuzzy systems, as selecting good design (number of hidden layers, number of neurons in a layer), choosing best learning algorithm (BP, LM, etc.) or selecting proper rule base is very crucial for them; because various structures can lead to very different and non-optimal results [27], [28], [30]. Validation is very important for NNs, because once they are trained on training set, they should provide good forecasts for out-of-sample test data too and for this, some performance evaluation error criteria (like MAE, MSE, RMSE, MAPE or COD) are used. Choosing a particular criterion from various

options is very difficult because use of different criterion lead to different results every time evident from [30], [48], and so, none of them can be used universally. This indicates the need of a universal and robust criterion for evaluating the performance of system during training and testing processes. B. Comparative Evaluation Evaluation of various methods based on their time-scales is presented in this section. 1) Very short-term forecast For this interval, there are not many papers are available. One of the very short-term forecasting methods reviewed is based on hybrid modeling (ANFIS system) [3], [10]. It provides significant improvement over the persistence method for 5 to 15 min-ahead period. Another is based on combination of two statistical approaches (time-series and filtering) that gave good COD for 1 to 5s ahead forecasting but its practicality in restructured electricity markets is doubtful as there are not markets clearing at such short period. However, it can be used for regulation purposes as required to respond to fault tackling, quick load changes, etc. So, it seems ANFIS model is the better one at present. Also, there is a requirement for models producing 30 min-ahead forecasts as in many deregulated markets settlements are done every 30 min. 2) Short-term forecast For this term, a large variety of methods is available. There is no clear winner but some better models are pointed out. Abilities of time-series models (mostly ARMA based) differ when applied to different time periods. These are usually superseded by other techniques like NNs, hybrid approaches. This should not necessarily be the case. A few good methods, like Mycielski algorithm [31], GM [33], [34], perform very well. Also, hybridizing alternative time-series models [25], inclusion of other variables like wind direction, pressure along with wind speed or hybridization of time-series model with some newer approaches, like wavelet transforms [19], GP [29] or Bayesian estimation, have shown good results. Among the NN models, those including atmospheric variables perform better than the ones just having historical data [26]. Also, it is seen that complex NN structures like RNNs, Adaline networks have advantage over conventional ones [30], [44], [45]. However, this may not be always true as many times use of simple FNNs gave good results [28], [37], [47]. In the literature, hybrid concepts are given a good emphasis. Hybridization of NNs with other methods produces very good forecasts as shown by ANFIS for very short-term forecast. A popular hybrid method is evaluating spatial correlations between various wind sites (one local and few neighboring) using NNs or fuzzy logic [17], [18]. This technique is very robust and gives one of the best results. There are few new methods like complex statistics, but there is still lot more to do and prove for these new methods. So, although there is no clear winner; NNs or hybridized NNs give overall good results for short-term forecasting. 3) Medium- and Long-term forecast Medium- and long-term forecasting use very similar

techniques and hence, methods used for them have been evaluated together. Due to long ahead-forecasting duration, simple models dont remain as accurate as they do for shortterm prediction and so, use of NWP and hybrid NWP models is very prevalent as they give best results for these time-scales. Various NWP systems such as Global Forecasting System, MM5, Prediktor, HIRLAM have been developed, but just NWP models cannot provide sufficient downscaling for a particular wind park in a particular location [20], [46]. That is why hybrid NWP is developed. In hybrid NWPs, forecasting results from NWPs are used as preliminary results and they are fed as input along with generation data from utility to NN or fuzzy structures or other hybrid structures like ANFIS or their combination with other statistical time-series methods like ARMA models, which provide enough downscaling and very good forecasts for a specific wind farm [20], [21], [40], [46]. One important observation for these time-scales is the use of non-Gaussian error distribution function [42] and using entropy based criteria for testing the training performance of NNs [48]. Also, online training of model seemed more effective than offline training [48]. These concepts have been proposed recently showing some test cases but have not been adopted fully in any of the techniques yet. Their inclusion would take wind forecasting to a new level. VII. CONCLUSIONS This paper presented a review on forecasting of wind speed and generated power considering different time scales. Several forecasting models were discussed and a lot of researches on the models, which have their own characteristics, were presented. The major focus was on emphasizing the diversity of various forecasting methods available and also on providing a comparison of present mechanisms to determine the best available. REFERENCES
[1] Y-K Wu, and J-S Hong, A literature review of wind forecasting [2] [3] [4]
technology in the world, IEEE Power Tech 2007, Lausanne , pp. 504509, 1-5 July 2007. H. Lund, Large-scale integration of wind power into different energy systems, Energy, vol. 30, no. 13, pp. 2402-2412, Oct. 2005. M. Negnevitsky, P. Johnson, and S. Santoso, Short term wind power forecasting using hybrid intelligent systems, IEEE Power Engineering Society General Meeting 2007, pp.1-4, 24-28 June 2007. A. Fabbri, T. G. S. Roman, J. R. Abbad, and V. H. M. Quezada, Assessment of the cost associated with wind generation prediction errors in a liberalized electricity market, IEEE Trans. Power Syst., vol. 20, no.3, pp. 1440- 1446, Aug. 2005. P. Pinson, C. Chevallier, and G. N. Kariniotakis, Trading wind generation from short-term probabilistic forecasts of wind power, IEEE Trans. Power Syst., vol. 22, no.3, pp.1148-1156, Aug. 2007. S. J. Watson, L. Landberg, and J. A. Halliday, Application of wind speed forecasting to the integration of wind energy into a large scale power system, IEE Proc. - Gener. Transm. Distrib., vol. 141, no. 4, pp. 357-362, July 1994. N. R. Ullah, K. Bhattacharya, and T. Thiringer, Wind farms as reactive power ancillary service providers technical and economic issues, IEEE Trans. Energy Convers., vol. 24, no. 3, pp. 661-672, Sept. 2009. M. Lei, L. Shiyan, J. Chuanwen, Liu Hongling, and Z. Yan, A review on the forecasting of wind speed and generated power, Renewable and Sustainable Energy Reviews, vol. 13, no. 4, pp. 915-920, May 2009.

[5] [6]

[7] [8]

[9] A. Costa, A. Crespo, J. Navarro, G. Lizcano, H. Madsen, and E. Feitosa,


A review on the young history of the wind power short-term prediction, Renewable and Sustainable Energy Reviews, vol. 12, no. 6, pp. 1725-1744, Aug. 2008. C. W. Potter, and M. Negnevitsky, Very short-term wind forecasting for Tasmanian power generation, IEEE Trans. Power Syst., vol. 21, no. 2, pp. 965- 972, May 2006. M. Milligan, M. Schwartz, and Y. Wan, Statistical wind power forecasting models: Results for U.S. wind farms, in Proc. Windpower, Austin, TX, May 1821, 2003, NREL/CP-500-33 956 Rep. M. Lange, and U. Focken, New developments in wind energy forecasting, IEEE Power and Energy Society General Meeting 2008 Conversion and Delivery of Electrical Energy in the 21st Century, pp. 1-8, 20-24 July 2008. B. Candy, S. J. English, and S. J. Keogh, A Comparison of the impact of QuikScat and WindSat wind vector products on met office analyses and forecasts, IEEE Trans. Geosci. Remote Sens., vol. 47, no.6, pp. 1632-1640, June 2009. E. Cadenas, and W. Rivera, Wind speed forecasting in the South Coast of Oaxaca, Mexico, Renewable Energy, vol. 32, no. 12, pp. 2116-2128, Oct. 2007. E. Cadenas, O.A. Jaramillo, and W. Rivera, Analysis and forecasting of wind velocity in chetumal, quintana roo, using the single exponential smoothing method, Renewable Energy, vol. 35, no. 5, pp. 925-930, May 2010. P. S. Chang, and L. Li, Ocean surface wind speed and direction retrievals from the SSM/I, IEEE Trans. Geosci. Remote Sens., vol. 36, no. 6, pp. 1866-1871, Nov 1998. M. C. Alexiadis, P. S. Dokopoulos, and H. S. Sahsamanoglou, Wind speed and power forecasting based on spatial correlation models, IEEE Trans. Energy Convers., vol. 14, no. 3, pp. 836-842, Sept. 1999. I. G. Damousis, M. C. Alexiadis, J. B. Theocharis, and P. S. Dokopoulos, A fuzzy model for wind speed prediction and power generation in wind parks using spatial correlation, IEEE Trans. Energy Convers., vol. 19, no. 2, pp. 352- 361, June 2004. H. Liu, H-Q Tian, C. Chen, and Y. Li, A hybrid statistical method to predict wind speed and wind power, Renewable Energy, vol. 35, no. 8, pp. 1857-1861, Aug. 2010. G. Sideratos, and N. D. Hatziargyriou, An Advanced Statistical Method for Wind Power Forecasting, IEEE Trans. Power Syst., vol. 22, no. 1, pp. 258-265, Feb. 2007. I. J. Ramirez-Rosado, L. A. Fernandez-Jimenez, C. Monteiro, J. Sousa, and R. Bessa, Comparison of two new short-term wind-power forecasting systems, Renewable Energy, Vol.34, Issue 7, pp. 18481854, Jul. 2009. P. Flores, A. Tapia, and G. Tapia, Application of a control algorithm for wind speed prediction and active power generation, Renewable Energy, vol. 30, no. 4, pp. 523-536, Apr. 2005. G. H. Riahy, and M. Abedi, Short term wind speed forecasting for wind turbine applications using linear prediction method, Renewable Energy, vol. 33, no. 1, pp. 35-41, January 2008. M. S. Miranda, and R. W. Dunn, One-hour-ahead wind speed prediction using a Bayesian methodology, IEEE Power Engineering Society General Meeting 2006, pp. 1-6. A. R. Garcia, and E. De-La-Torre-Vega, A Statistical wind power forecasting system A Mexican wind-farm case study, European Wind Energy Conference & Exhibition EWEC Parc Chanot, Marseille, France, March 2009. E. Panteri, and S. Papathanassiou, Evaluation of two simple wind power forecasting models, in Proc. of the European Wind Energy Conf., National Technical University of Athens (NTUA), School of Electrical and Computer, 2008. G. N. Kariniotakis, G. S. Stavrakakis, and E. F. Nogaret, Wind power forecasting using advanced neural networks models, IEEE Trans. Energy Convers., vol. 11, no. 4, pp.762-767, Dec. 1996. E. Cadenas, and W. Rivera, Short term wind speed forecasting in La Venta, Oaxaca, Mexico, using artificial neural networks, Renewable Energy, vol. 34, no. 1, pp. 274-278, Jan. 2009. H. Mori, and E. Kurata, Application of Gaussian Process to wind speed forecasting for wind power generation, IEEE International Conf. Sustainable Energy Technologies (ICSET), pp. 956-959, 24-27 Nov. 2008.

[30] G. Li, and J. Shi, On comparing three artificial neural networks for [31] [32] [33] [34] [35] [36] [37]
wind speed forecasting, Applied Energy, vol. 87, no. 7, pp. 2313-2320, Jul. 2010. F. O. Hocaoglu, M. Fidan, and O. N. Gerek, Mycielski approach for wind speed prediction, Energy Conversion and Management, vol. 50, no. 6, pp. 1436-1443, June 2009. M. Monfared, H. Rastegar, and H. M. Kojabadi, A new strategy for wind speed forecasting using artificial intelligent methods, Renewable Energy, vol. 34, Issue 3, pp. 845-848, Mar. 2009. T. H. M. El-Fouly, E. F. El-Saadany, and M. M. A. Salama, Grey predictor for wind energy conversion systems output power prediction, IEEE Trans. Power Syst., vol. 21, no. 3, pp. 1450-1452, Aug. 2006. T. H. M. El-Fouly, E. F. El-Saadany, and M. M. A. Salama, Improved Grey predictor rolling models for wind power prediction, IET Gener. Trans. & Dist., vol.1, no.6, pp. 928-937, Nov. 2007. D. P. Mandic, S. Javidi, S.L. Goh, A. Kuh, and K. Aihara, Complexvalued prediction of wind profile using augmented complex statistics, Renewable Energy, vol. 34, no. 1, pp. 196-201, Jan. 2009. A. A. Khan, and M. Shahidehpour, One day ahead wind speed forecasting using wavelets, Power Systems Conference and Exposition, 2009, PSCE '09. IEEE/PES, pp. 1-5, 15-18 March 2009. J. P. S. Catalao, H. M. I. Pousinho, and V. M. F. Mendes, An artificial neural network approach for short-term wind power forecasting in Portugal, 15th International Conference on Intelligent System Applications to Power Systems, 2009, ISAP '09, pp.1-5, 8-12 Nov. 2009. H. Madsen, G. Kariniotakis, H. A. Nielsen, T. S. Nielsen, and P. Pinson, "A protocol for standardising the performance evaluation of short-term wind power prediction models," in Proc. Global Wind-Power Conf. (CD-ROM), Chicago, IL, Mar. 28-31, 2004. R. G. Kavasseri, and K. Seetharaman, Day-ahead wind speed forecasting using f-ARIMA models, Renewable Energy, vol. 34, no. 5, pp. 1388-1393, May 2009. L. Lazic, G. Pejanovic, and M. Zivkovic, Wind forecasts for wind power generation using the Eta model, Renewable Energy, vol. 35, no. 6, pp. 1236-1243, June 2010. S. Fan, J. R. Liao, R. Yokoyama, L. Chen, and W-J Lee, Forecasting the wind generation using a two-stage network based on meteorological information, IEEE Trans. Energy Convers., vol.24, no.2, pp.474-482, June 2009. H. Bludszuweit, J. A. Dominguez-Navarro, and A. Llombart, Statistical analysis of wind power forecast error, IEEE Trans. Power Syst., vol. 23, no. 3, pp. 983-991, Aug. 2008. J. W. Taylor, P. E. McSharry, and R. Buizza, Wind power density forecasting using ensemble predictions and time series models, IEEE Trans. Energy Convers., vol. 24, no. 3, pp. 775-782, Sept. 2009. T. G. Barbounis, J. B. Theocharis, M. C. Alexiadis, and P. S. Dokopoulos, Long-term wind speed and power forecasting using local recurrent neural network models, IEEE Trans. Energy Convers., vol. 21, no. 1, pp. 273- 284, March 2006. T. Senjyu, A. Yona, N. Urasaki, and T. Funabashi, Application of recurrent neural network to long-term-ahead generating power Forecasting for Wind Power Generator, IEEE PES Power Systems Conference & Exposition (PSCE) 2006, pp. 1260-1265, Oct. 29 -Nov. 1, 2006. S. Salcedo-Sanz, A. M. Perez-Bellido, E. G. Ortiz-Garcia, A. PortillaFigueras, L. Prieto, and D. Paredes, Hybridizing the fifth generation mesoscale model with artificial neural networks for short-term wind speed prediction, Renewable Energy, vol. 34, no. 6, pp. 1451-1457, June 2009. D. A. Fadare, The application of artificial neural networks to mapping of wind speed profile for energy application in Nigeria, Applied Energy, vol.87, no. 3, pp. 934-942, Mar. 2010. R. J. Bessa, V. Miranda, and J. Gama, Entropy and correntropy against minimum square error in offline and online three-day ahead wind power forecasting, IEEE Trans. Power Syst., vol. 24, no. 4, pp. 1657-1666, Nov. 2009. H. S. Hippert, C. E. Pedreira and R. C. Souza, Neural networks for short-term load forecasting: a review and evaluation, IEEE Trans. Power Syst., vol. 16, no. 1, pp. 44-55, Feb. 2001.

[10] [11] [12]

[13]

[14] [15]

[16] [17] [18]

[38]

[39] [40] [41]

[19] [20] [21]

[42] [43] [44]

[22] [23] [24] [25]

[45]

[46]

[26]

[47] [48]

[27] [28] [29]

[49]

Das könnte Ihnen auch gefallen