1. Introduction
Global climatic conditions have changed rapidly over the last decades. The continuous increase of energy needs all over the world, as well as the use of limited reserves of traditional energy resources (coal, oil, and natural gas), have turned the interest of researchers towards renewable energy. One of the most important and widely used renewable resources is wind power, thanks to its widely distributed nature [
1]. Considering the advance in technology and research, wind power has become an indispensable part of the global energy system and it could be possible that in the future it could largely replace conventional energy resources used for power generation.
Because of the wind’s stochastic nature and intermittence, the increased penetration of wind power has created many challenges in the operation and planning of power systems worldwide [
1]. To deal with these challenges, it has been necessary to develop wind power forecasting models and methods with increased accuracy. As a result, wind power forecasting has been researched and developed over the past decades in order to deal with the challenges that arose with the rapid increase in the use of wind power in the power systems worldwide. Forecasting models not only forecast wind power, but also help stabilize power systems and organize electricity markets [
2].
Deterministic forecasting models have been used over the last decades in wind power generation and play an important role in the daily operation of power systems. Given a set of input data, deterministic forecasting models are able to provide the user with a single-valued expectation series of the wind power output. Depending on the evaluation of the model used and its errors, the user is able to use the model’s results to estimate the closest possible output of wind power generation. For this, numerous deterministic forecasting models have been proposed and developed to predict as accurately as possible the wind power output. Naturally, such models, thanks to the development of technology, are still improving in order to provide better forecasts [
3].
Over the last decade, probabilistic forecasting has been the center of attention for researchers since, unlike deterministic forecasting, probabilistic forecasts give important information over the uncertainty of the forecasts. While deterministic methods give single-valued results of wind power generation, probabilistic methods give a wider view of possible wind power outputs since the output of such models could be quantiles, prediction intervals (PIs), and distributions. In this way, the user has a better view of the possible forecast compared to the single-value output of a conventional deterministic model. As a result, it could be possible in the future that probabilistic forecasts could be used effectively in decision-making problems, transforming various decision-making activities to probabilistic, such as wind power trading in electricity markets [
4], optimal power flow [
5], and unit commitment [
6].
Evaluation is a very important aspect of wind power forecasting. Evaluating proposed forecasting models allows a constant comparison between different models and consequently their constant development. Evaluation is important in both deterministic and probabilistic forecasts. In deterministic forecasting, simple comparative measures have been used over the years to evaluate the performance of the forecasting models. However, evaluating probabilistic forecasts is more complicated than evaluating point predictions. While in the point forecasts the evaluation is based on the deviation between predicted and measured power values, the same is not possible in probabilistic forecasting since such a comparison is not possible directly. This is why defining a framework to evaluate probabilistic forecasts has been researched a lot over the past years.
With the increased importance of wind power in energy systems, more and more researchers not only intend to create new advanced models, but also observe the function of existing models in various cases. Recently, review works have studied deterministic and probabilistic forecasting models in order to comprehend the methodologies used in wind power forecasting as well as determine possible development possibilities in the future. More specifically, [
7] divided the deterministic forecasting methods reviewed in four categories and presented their different characteristics. In [
8], an in-depth review of wind power forecasting methods was presented as well as an overview of benchmark techniques and uncertainty analysis. The review works [
9,
10] classified the reviewed deterministic forecasting works from the perspective of forecasting horizons and time scales. In [
11], various works were reviewed and the forecasting accuracy of the models based on the variable factors used in the forecast was discussed. The work [
12] reviewed state-of-the-art probabilistic forecasting models and presented an overall framework of probabilistic forecasting evaluation. The work [
13] presented the fundamental concepts of state-of-the-art probabilistic methodologies. The work [
14] focused on the principals and features of state-of-the-art wind power forecasting uncertainty analysis.
The above bibliography review shows that the majority of the review papers are focused on presenting the state-of-the-art methodologies of either deterministic or probabilistic wind power forecasting. However, there is no review work focused on presenting evaluation results in order to propose a comparative knowledge between different methodologies in specific conditions.
The contributions of this review paper are manifold:
It offers a unique and wider view by reviewing the state of the art in both the deterministic and probabilistic wind power forecasting methodologies and by identifying their advantages and disadvantages.
It provides comparative results among the models of the reviewed works, based on evaluation measures.
It proposes future research goals of deterministic and probabilistic forecasting models, not only to improve the methodologies used in forecasting, but also to make both deterministic and probabilistic forecasting models more useful and helpful in energy markets and power systems.
It aims to help researchers in having a view of possible expectations in research cases corresponding to the ones compared in the paper.
The structure of the paper is as follows.
Section 2 analyzes the evaluation methods used for assessing the deterministic and probabilistic forecasting methodologies.
Section 3 analyzes state-of-the-art wind power deterministic forecasting methodologies.
Section 4 describes state-of-the-art wind power probabilistic forecasting methodologies.
Section 5 analyzes recent wind power forecasting methodologies.
Section 6 provides the advantages and disadvantages of using wind power deterministic and probabilistic forecasting methods.
Section 7 identifies the core contribution of the reviewed works for wind power deterministic and probabilistic forecasting methods.
Section 8 presents the comparative results between the methodologies compared in the reviewed works.
Section 9 proposes future research directions and possibilities.
Section 10 summarizes the main findings and concludes the paper.
2. Evaluation of Wind Power Forecasts
Evaluating a forecasting model is of core importance to comprehending its quality and accuracy. The continuous need to improve existing forecasting models and develop new ones, calls for a way to compare such models concerning the accuracy of their results. In deterministic forecasting as well as in probabilistic forecasting, different ways of comparing and evaluating different models have been used throughout the years.
2.1. Deterministic Forecasting
There are different ways to evaluate deterministic models. Evaluation of the models is an important procedure, since it allows the comparison between different models, highlights possible problems in the methodologies used, and guides possible improvements. Furthermore, the evaluation of deterministic models is quite easy to use and understand, as it consists of simple error metrics that can be easily compared between different models. Such error metrics include:
As described in (1),
MAE is the average value of the absolute error of the
N forecasted error values
en.
MSE is the average value of the squared error, which is the squared difference between the actual and predicted values. It can be described as:
As described in (3),
RMSE represents the squared root of the quadratic mean of the difference between the actual and predicted values:
2.2. Probabilistic Forecasting
Evaluating wind power probabilistic forecasts is not that simple. While in deterministic forecasting the evaluation takes place by comparing simple error metrics, in probabilistic forecasting the evaluation is a more complicated process. Considering the fact that wind power generation itself is a complex process, the uncertainty of the probabilistic forecasting models is influenced by a large number of external factors.
The properties needed to evaluate probabilistic predictions are reliability, sharpness, resolution and skill score. Reliability is usually the first parameter that should be evaluated. Since non-parametric probabilistic predictions include a single quantile forecast, or a collection of quantile forecasts, the evaluation of their reliability is based on verifying the reliability of each one of those quantile forecasts. A critical measure that is widely used is the Prediction Interval Coverage Probability (
PICP):
where
Ntest is the number of the test samples used, and
ci is the
PICP’s indicator that is defined as:
where
ti is the measured target and
Lt and
Ut are the lower and upper bounds of the prediction interval, respectively.
PICP is directly connected and compared to the Prediction Interval Nominal Coverage (
PINC) percentage, presented in (6):
where (1 −
α) expresses the nominal coverage probability of the prediction intervals. The closer the
PICP is to the
PINC the more reliable is the PI. Another simple metric usually used is the Average Coverage Error (
ACE):
The closer the ACE value is to zero, the higher the reliability of the PI.
Sharpness refers to the width of the prediction intervals derived from the forecasting process. A narrower PI is generally preferred over a wider one, since it narrows down the possible outcomes of the forecast and therefore offers better information to the user, facilitating decision making. A common measure used to estimate and control the PI’s width is the PI Normalized Average Width (
PINAW), defined as:
where
R is the range of the underlying targets that were used for normalizing PIs.
The PI Normalized Root-mean-square Width (
PINRW) is also used and is defined as:
Another measure that is widely used is the Coverage Width Criterion (
CWC), which simultaneously defines both the reliability and the sharpness of a PI.
CWC is calculated by (10):
where
μ is the confidence level and it usually equals 1 −
α;
η is a penalty coefficient used to increase the difference between
PICP and
μ whenever
PICP is smaller than
μ; and
γ(
PICP,
μ) is defined as follows:
The resolution refers to the ability of a forecasting model to generate different probabilistic information according to the forecast conditions and to provide reliable and accurate forecasting distributions.
Skill score is another important evaluation property that allows the comparison between different predictive approaches, since a higher overall skill score suggests a higher skill of the proposed probabilistic forecasting model over other predictive models.
The use of the above evaluation tools (a) allows the evaluation of a specific wind power probabilistic forecasting (WPPF) model; (b) allows the comparison of a model with other WPPF models; and (c) may identify possible problems in aWPPF model that is under development and evaluation. As a result, WPPF models can keep improving their accuracy and performance and also become simpler for the user to understand.
5. Recent Wind Power Forecasting Methodologies
With the advances in technology and computer capabilities, new methodologies are being used in order to improve the accuracy of the forecasting models. Such methodologies include deep learning, reinforcement learning, as well as improvement in spatio-temporal forecasting.
The work [
85] proposed a hybrid model with three different neural networks for the predictions. A reinforcement method based on ensemble learning was used to improve the model’s accuracy as well as the deep networks’ efficiency. When compared to various models and state-of-the-art methodologies, the proposed model managed to outperform the other methods, thanks to the combination of the Empirical Wavelet Transform decomposition (EWT), the ensemble learning method, and deep networks.
The work [
86] proposed Markov-chain-based stochastic models for short-term distributional forecasts and wind farm generation forecasts. The distributional forecast could be further used in problems of unit commitment and economic dispatch in order to transform them into problems studied under a Markov-chain-based stochastic framework. It was also observed that the spatial dynamics of a single wind farm could differ significantly. Concerning the distributional forecasts, the proposed model managed to outperform two different high-order autoregressive (AR) models. As for the point forecasts, the proposed model managed to outperform both AR models, as well as the persistence model. However, it was observed that for the proposed model’s accuracy to increase, the model itself had to become more complex.
In [
87], a bi-level CNN model was used in order to improve the accuracy of wind power forecasts, thanks to the CNN’s deep feature extraction capabilities. The methodology used also combined Variational Mode Decomposition (VMD) and Phase Space Reconstruction (PSR) for data preprocessing and PSO algorithm for the final optimization. The proposed model was compared to the persistence method, a single CNN model and a VPCB (VMD+PSR+CNN+BPNN) model and managed to outperform them in deterministic (based on NMAE, NRMSE and MAPE metrics) as well as in probabilistic forecasting (based on
PINAW measure).
In [
88], a Compressive Spatio-Temporal (CST) method was proposed for wind power forecasting along with WT in order to improve the model’s performance. The main focus was deterministic wind speed forecasting results, especially for longer prediction horizons. However, it was noted that the model could be also used for wind power deterministic and probabilistic forecasting.
In the work [
89], a deep mixture network was designed in order to directly construct PDFs from data. CNN and gated recurrent unit (GRU) were also used in order to learn spatio-temporal features of high volatility wind speed time series. The proposed model was tested for two different datasets. In both data sets, the proposed model managed to outperform several mixture models in terms of performance. Based on the continuously ranked probability score (CRPS) and cross-entropy (CE) metrics, the proposed model gave better results, concerning its sharpness, accuracy and reliability.
In the work [
90], a novel WPPF artificial intelligence model is proposed, where spiking neural networks (SNN) along with group search optimizer (GSO) are used. The SNN is used for the training of the wind power data and the GSO was implemented to optimize the parameters of the SNN. The proposed methodology was compared to other benchmark models (BPNN, SVM, ELM). Based on the evaluation metrics (
ACE, IS), the proposed methodology had an overall higher performance and accuracy. Moreover, the probability coverage of the proposed method improved by 72.0%, 54.9% and 51.3% when compared to the BPNN, SVM and ELM, respectively. Furthermore, the computational cost is sufficiently low, where the SNN was able to train 30,000 samples in less than 3 s.
In the work [
91], a novel multi-model combination (MMC) method was proposed that combined different forecasting models. A two-step optimization methodology based on the expectation maximizing (EM) algorithm was used to estimate weights of member models. In terms of reliability, the proposed MMC model with further optimization (FO) outperformed the other state-of-the-art models it was compared to. Furthermore, in order to estimate the comprehensive performance according to the CPRS metric, the proposed MMC+EM+FO model showed improved calibration compared to other models. As a result, the proposed model was tested in different wind farms and was proposed to be used for real time system operation problems.
The work [
92] proposed a deep learning model for wind power forecasting. The model consisted of two different SVMs which had their own specific loss function. Each SVM model was used in a way that it improved its position according to the reaction of the other one. While the first support vector tried to “deceive” the second one, the second support vector aimed to improve its training process in order to avoid wrong decisions. Moreover, a modified flower pollination (MFP) algorithm was developed to optimize and adjust the parameters of the support vectors. The LUBE model was used to acquire the PIs. The proposed model using the MFP algorithm was compared to other benchmark optimizers, such as the PSO and the GA via the Confidence Level Index (CLI) and the Average Bandwidth Index (ABI) metrics. To improve both metrics, a fuzzy min-max solution was used were the objectives were to maximize the CLI and minimize the ABI. Based on the results, the proposed MFP model outperformed the PSO and GA ones, as it gave the best solution for the CLI and ABI metrics, with values of 94.76356 and 29.86350, respectively, while the respective values for the PSO were 86.03972 and 34.03482, and for the GA they were 91.16583 and 31.11378.
In [
93], an Improved Deep Mixture Density Network (IDMDN) was proposed in order to process the wind power probabilistic forecasting of multiple wind farms. A beta kernel function was also used in order to avoid the density leakage problem. Due to its end-to-end architecture, the proposed model was data-adaptive and could extend to other regions. The proposed model was used for both deterministic and probabilistic forecasting. Concerning the deterministic forecasts, the model was used for forecasting in seven different wind farms. Based on the NRMSE error metric, when compared to various different models, the proposed model outperformed the others in five of the seven wind farms, while based on the NMAE error metric, it outperformed the other models in four of the seven wind farms. Concerning the deterministic forecasts, the
ACE,
PINAW, IS and CRPS metrics were used for the comparison. The model was compared to the Deep Belief Network (DBN), the LSTM network and the Gradient Boosting Machine (GBM) models. In terms of the
ACE,
PINAW and CRPS metrics, the proposed IDMDN model gave better results, while in terms of the IS metric, the GBM model was slightly better.
The work [
94] proposed a Time Warping Invariant Echo State Network (TWIESN) based on an advanced reservoir computing framework for the WPF process. Furthermore, a Multi-Objective Grey Enhanced Wolf Algorithm (MOEGWA) was developed for the optimization in order to improve the stability and accuracy of the model. Moreover, regressional ReliefF (RReliefF) algorithm along with the Granger Causal Relation Test (GCRT) were developed in order to estimate and select the most appropriate candidates for TWIESN from original input features. The proposed model was a RreliefF-GCRT-MOEGWA-TWIESN model, and it was used for deterministic and probabilistic forecasting results. Considering the deterministic forecasting, the proposed model was compared to various benchmark models which, in terms of MAPE,
MAE,
RMSE, NMSE error metrics managed to outperform. Considering the probabilistic forecasting, the proposed model was constructed using the Gaussian distribution and the T location-scale (TLS) distribution. The proposed methodology was considered the one with the TLS distribution. When compared to the Gaussian distribution as well as the quantile regression (QR), for different
PINCs, based on
PICP,
PINAW, AWD and
ACE metrics, the proposed methodology outperformed both the QR and the Gaussian distribution.
8. Comparative Results of Reviewed Works
The comparative results of the reviewed works in deterministic and probabilistic forecasting are provided in
Table 5 and
Table 6, respectively.
In
Table 5, numerous reviewed works are presented, based on deterministic forecasting models. Each work proposed a specific model for wind power forecasting that was later compared to other benchmark methodologies in order to prove its efficiency. The compared methodologies as well as the parameters of evaluation used for the comparison are also presented in
Table 5.
In the work [
24], an ARIMA-ARCH model was proposed for short term wind forecasting. The proposed model was compared to a single-ARIMA methodology. The MRE metric was used to compare the two methodologies. The ARIMA-ARCH model had an MRE of 11.2% while the single-ARIMA one had an MRE of 17.4, showing the improvement of the proposed model in the forecasting error.
In the work [
25], a fractional-ARIMA (f-ARIMA) model was proposed. The performance of the f-ARIMA model was compared to that of the persistence method and the single-ARIMA method, in terms of the error metrics of the DME, the variance (
σ2) and the square root of the forecast mean square error. Based on the DME, the proposed model had a value of 33.18, while the persistence and the single-ARIMA models had a value of 45.2 and 144.92, respectively. The proposed model was found superior in the other error metrics too.
The work [
20] compared the ARMA model with ANN models and the persistence model in three different cases for short-term wind power forecasting. Based on the
MAE,
RMSE and MRE error metrics, the ARMA model outperformed the other models in all three cases. However, it should be noted that the ARMA model shown higher processing time in all three cases.
In [
21], an ARMA-Pattern Matching model was proposed for short-term and ultra-short-term wind power forecasting. The proposed model was compared to the traditional ARMA model based on the relative tolerance. According to the results, the traditional ARMA model gives sufficient results for 0–1 h forecasting. However, with the increase in the forecast time (1–6 h), the pattern-matching model gave greatly more accurate results.
The work [
22] aimed to investigate the efficiency of various ARMA models in short-term wind forecasting. The different ARMA models were compared to the persistence model in order to prove their forecasting ability. While in their majority, the ARMA models surpassed the persistence model, the ARMA (0, 36) beat the persistence method for eight 10-min periods ahead. More specifically, based on the
RMSE metric, the ARMA (0, 36) had a 16%
RMSE improvement for one period ahead, an 8% improvement for three periods ahead and 7.5% improvement for four periods ahead.
In the work [
29], the performance of three different ANN models (FFBP, RBF, ADALINE) was researched for 1-h ahead wind forecasting. The three ANNs were tested in two different sites, while the error metrics used for the comparison of the models were the
MAE, MAPE and
RMSE. In the first case, in terms of MAPE, the ADALINE model outperformed the FFBP and RBF models by 4.8% and 14.0%, respectively. In terms of
MAE and
RMSE, the FFBP model outperformed the other ANNs with values of 0.951 and 1.254, respectively. In the second case, the RBF model surpassed the other methodologies in all error metrics.
The work [
30] investigated the accuracy of different ANN models for short-term wind forecasting. The study used four different ANN configurations, ANN1 (3 layers: 3 input, 3 hidden layers, 1 output), ANN2 (3 layers: 3 input, 2 hidden layers, 1 output), ANN3 (2 layers: 3 input, 1 output) and ANN4 (2 layers: 2 input, 1 output). Based on the error metrics’ results, in terms of
MSE, ANN4 had the lowest value with 0.0016 as well as in terms of
MAE with a value of 0.0399.
The work [
31] proposed an ANN approach for short-term wind power forecasting. The proposed methodology was compared to the persistence model, using the MAPE metric as the error metric. The average MAPE value of the proposed model was 7.26% while the average value of the persistence model was 19.05%, proving the superiority of the ANN approach.
In [
37], two WT-SVM models were tested for short-term wind forecasting. The WT-SVM methodologies were compared to the RBF-SVM model for different time scales. The results shown that based on the MRE and
RMSE, the WT-SVM model outperformed the RBF-SVM model in all time scales, i.e., for 1h, the MRE and
RMSE values for the Method 1 of the WT-SVM were 7.97 and 11.52, respectively, while for the RBF-SVM model the values were 10.50 and 16.19, respectively.
The work [
38] proposed a WT-SVM-GA hybrid model for short term wind forecasting. The proposed methodology was compared to the persistence model and the SVM-GA model without the implementation of the wavelet transform. The comparison between the models was based on the
MAE and MAPE error metrics. The results have shown that the proposed method significantly outperformed the other two methodologies. For example, based on the MAPE metric, the proposed model had a value of 14.79% while the persistence and the SVM-GA models had respective values of 22.64% and 17.8%.
In the work [
35], an SVM-based model was proposed for short-term and very-short-term wind power forecasting. The proposed model was compared to the persistence model and the RBF-NN model. Based on
MAE and MAPE error metrics the proposed model was superior in both short-term and very-short-term forecasting.
In the work [
39], an IDA-SVM-based model was proposed for short-term WPF. The proposed model was compared to different SVM models that used different optimization algorithms, as well as a simple Back Propagation Neural Network (BPNN) and the Gaussian Process Regression (GPR). Based on the NRMSE, NMAE, MAPE and R
2 error metrics, the proposed IDA-SVM model managed to outperform the other models in datasets from different seasons.
In [
40], a WTNN was proposed for short-term wind power forecasting. The proposed model was compared to the persistence model, the ARIMA model and simple NN models. In general, the average computational time of the proposed process was less than 10 s. Evaluating the different approaches, the proposed model had a MAPE value of 6.97% while the persistence model, the ARIMA and the NN models had a value of 19.05%, 19.34% and 7.26%, respectively.
The work [
41] tested two hybrid models, an ARIMA-ANN and an ARIMA-SVM for different time horizons. The models were compared to the single ARIMA, ANN and SVM models. The time horizons were 1, 3, 5, 7 and 9-step-ahead prediction times. The comparison was based on the
MAE and
RMSE error metrics. In terms of
MAE, the ARIMA-ANN model performed better for a 1-step time horizon, the single-ANN model performed better for 3-step and 7-step time horizons, and the single-SVM model had better performance for 5-step and 9-step ahead time horizons. In terms of
RMSE, the ARIMA-ANN model outperformed the other models for 1-step and 7-step time horizons; the single-ANN model had better performance for 3-step and 5-step time horizons, while the single-SVM model gave better results for the 9-step-ahead time horizon.
The work [
42] proposed a hybrid EMD-SVM forecasting model for wind power forecasting. The proposed model was compared to the single-SVM model in order to prove the efficiency in using a hybrid-based SVM model. The RMS error metric was used to compare the two methodologies. According to the results, the RMS value of the proposed model was 15.63% while the RMS value of the single-SVM method was 35.40%.
In the work [
43], a hybrid ARIMA-ANN model was proposed for wind forecasting. The proposed methodology was compared to the single-ARIMA and single-ANN models, in three different sites, with different datasets. The ME,
MAE and
MSE error metrics were used to compare the three different models. The results showed that, in all the three different sites, the proposed hybrid model had significantly better performance, as the results of all error metrics were lower than the single-ARIMA and single-ANN models.
In the work [
45], a hybrid LSTM neural network along with the Wavelet Decomposition (WD-LSTM) was used for wind power forecasting. The proposed model was compared to the BMA-EL, the MRMLE-AMS and the SVR-IDA. The proposed model managed to outperform the others, with a MAPE value of 5.831, while the BMA-EL had 22.328, the MRMLE-AMD had 20.624 and the SVR-IDA had 15.679.
In [
46], the accuracy of two different hybrid models was tested based on ANNs, PSO and GA. The two models were the PSO-PSO-ANN and the GA-PSO-ANN. The proposed models were compared to the single PSO-ANN model and the Adam-ANN model. Based on the MAPE and
MSE error metrics, both proposed models gave better results than the PSO-ANN and the Adam-ANN models, while their values in both MAPE and
MSE were similar.
9. Discussion and Future Research
Developing wind power forecasts has become really important in recent decades. Every year, there is higher and higher penetration of wind power in global electricity systems due to the necessity to use more renewable sources of energy. Therefore, various forecasting models have been developed and researched in order to properly use wind power as efficiently as possible. However, improving said models and researching new ones in order to improve the accuracy of wind power forecasts is still a matter of crucial importance.
Deterministic forecasts have been the center of attention in wind power forecasting for many years. Many different models have been developed in order to give accurate results for different problems, i.e., different time scales, different topographic conditions, and different datasets. A summary of those models and their basic features is presented in
Table 7. Nowadays, point forecasts play an important role not only in the operation of the power system but also in the operation of the electricity market. Accurate point forecasts are vital in order to maintain the stability of power system. As a result, improving the accuracy of existing forecasting models as well as developing new ones should be the main focus of researchers in the future.
Further development of AI forecasting models should be a future research area. Improving the training algorithms of those models could improve their accuracy significantly. Moreover, input data processing should be further researched for different cases as it can greatly affect the forecasting accuracy of the models. Furthermore, with the development of the possibilities of technology, novel AI models could be developed in order to give better results overall.
Another important aspect of wind power forecasts that should be further developed is NWP models. Such models are usually used as input data in many forecasting models as they use atmospheric data in order to make accurate wind speed value predictions. As a result, they are a useful tool for various models and improving their performance could result in improving the quality of input data used in forecasting.
In recent years, researchers have combined different types of deterministic models in order to improve the results of wind power forecasts [
40,
41,
42,
43]. Such hybrid approaches have managed to combine physical and statistical models effectively in different ways and have provided encouraging results. By considering the advantages of the combined methodologies, the whole process of a hybrid model could surpass conventional models not only in accuracy, but also in computational cost. However, despite the promising results hybrid methods have had so far, further research is needed not only to study existing hybrid models in different cases and conditions, but also to develop new ones that could offer much more accurate results.
On the other hand, WPPF models are still in an early stage of research. Point forecasts, which are widely used in wind power forecasting all over the world, aim to give single valued forecasts in order to make an accurate prediction. However, such predictions do not calculate the uncertainty part of a forecast, thus they do not offer any information over it. This is why in recent years, WPPF has been advancing rapidly and more and more researchers propose new models. Even though many of those models are still complicated to understand or evaluate, due to the early stage of research, recent WPPF models have managed to increase the performance and accuracy of uncertainty predictions. A classification of probabilistic forecasting models and their basic features is presented in
Table 8.
Probabilistic forecasting could play an important role in the future, not only in the operation of the energy market, but more importantly in decision making in power systems. Being able to give a wider view of a prediction, unlike spot forecasts, could prove to be an important tool in how decision makers use wind forecasts more efficiently in the future. However, for this to be possible, probabilistic forecasting models need to be further researched in the future in order to increase their accuracy, but also to make them more user friendly.
Recent research has been focused on spatio-temporal forecasting as interaction between wind parks is becoming necessary due to the fast integration of the wind power all over the world [
95,
96]. Spatio-temporal forecasting focuses on increasing the accuracy of the predictions via using information from different neighboring wind farms as predictors. Further developing spatio-temporal models could improve their accuracy significantly and make them an important tool in power systems. Furthermore, reinforcement learning along with deep learning have been more and more researched in recent years and further development in this domain seems promising and should be of main focus in the future.
Another aspect of WPPF that researchers have been recently focused on is ramp events. Ramp events pose a threat to power systems as wind power penetrates the global power system more and more every year [
97,
98]. Due to their dependency on many factors, such as weather conditions, different time scales, NWP input data accuracy, and multiple nearby locations, ramp forecasting models should be further developed in the future.
Forecasting has become an indispensable part of the stability and operation of the power systems all over the world. Deterministic models specifically have had and still have a significant effect on the operation and management of power systems and electricity markets. However, probabilistic models are still being developed and researched. Their importance in delivering results over the uncertainty part of a prediction shows that it is possible in the future, through novel probabilistic models and further research, for probabilistic forecasts to be used effectively in power systems. The methods for using uncertainty predictions as effectively as possible in the interest of power systems and energy markets could be an interesting field of research as well.
Further developing deterministic models as well as using probabilistic forecasts in other energy sources, i.e., solar power and geothermal power should be considered for future research. Achieving a more stable and organized power system should be the main focus of research in the future as it calls for a more complete understanding of the importance of the forecasts depending on the use of each renewable source. Since wind power is considered one of the most important renewable sources, it could be important to examine how its probabilistic models function with other energy sources and how the interactive forecasting results of different renewable types of energy influence the power systems and energy markets.
10. Conclusions
This paper presents a detailed overview of state-of-the-art wind power deterministic and probabilistic forecasting methodologies. Furthermore, it aims to provide a comparative overview through evaluation measures used for these methodologies in order to offer a view of possible expectations and outcomes of similar research.
Deterministic forecasting models have been widely used over the last decades in order to facilitate the wind penetration in power systems. Conventional time-series-based models, such as ARMA, ARIMA and Grey method as well as AI-based models, like ANN and SVM, have been the main focus of researchers all over the world. Said methodologies have managed, with the gradual advance of technology, to improve their point predictions to an efficient and useful level, in order to support the daily operation of electricity systems. Furthermore, hybrid methodologies proposed over the last years, have managed to further develop the possibilities of the conventional models via combining them in order to improve their accuracy and computational cost. The continuous technological development has also led new methodologies to be considered for wind power forecasting, like deep learning and reinforcement learning. Thanks to the evaluation processes that have been widely followed for many years, it was possible to compare the above models in various cases and estimate the most efficient methodology in each case individually.
The uncertainty of the predictions in wind power forecasting has also become a subject of great interest in recent studies. Various models have been proposed, like KDE, QR, LUBE, bootstrap, and ensemble-based methods. Those models have managed to give a wider view of possible outcomes of wind power predictions and thus provide an overall expectation of wind power values at a certain time point. Although probabilistic forecasting models are more difficult to evaluate, compared to deterministic models, specific metrics have been developed in order to satisfy this need. Reliability, sharpness, resolution and the overall skill score of a model have been used in order to compare different methodologies and estimate their efficiency. Based on those metrics, methodologies from different cases were compared to estimate the most efficient one.
The future research, proposed in this paper is mainly focused on improving the accuracy of existing deterministic forecasting models as well as developing more advanced probabilistic forecasting models in order to successfully implement them in power systems and energy markets management. Probabilistic forecasting models are still at an early stage of research. However, thanks to the continuous efforts to develop new models as well as the technological development, probabilistic forecasting has rapidly given accurate and efficient results in wind power forecasting. The use of both deterministic and probabilistic models in power systems as well as electricity markets, could be a possibility in the future as they could offer a more complete, stable and secure power system overall.