In early research on the prediction of renewable power generation, traditional machine learning methods were widely used, the most representative of which are support vector machines (SVMs) (Fonseca et al., 2012; Shi et al., 2012) and artificial neural networks (ANNs) (Izgi et al., 2012; Dumitru et al., 2016; Son et al., 2018). With the development of this research field, traditional machine learning methods have gradually been replaced by hybrid machine learning methods, ensemble learning methods, and deep learning methods.

Hybrid machine learning methods combine the advantages of more than two methods to achieve better prediction results. A key direction is to introduce intelligent optimization algorithms into traditional machine learning to improve the performance of the algorithms. The most commonly used hybrid methods in renewable energy generation forecasting have been based on support vector machines (SVMs) and extreme learning machines (ELMs). SVM is a nonlinear classifier based on the kernel method, which can avoid the disadvantages of overfitting (Olatomiwa et al., 2015). VanDeventer et al. (2019) introduced a genetic algorithm (GA) based on the support vector machine and demonstrated that the GASVM model can reduce the root mean square error (RMSE) by 98% in short-term PV power prediction compared to the conventional SVM model. Similarly, Xiao et al. (2022) proposed an SVM model based on gray wolf optimization (GWO-SVM), which can significantly reduce the RMSE in PV power prediction. ELM is a special single-layer feedforward neural network with fast learning ability and good generalization. Acikgoz et al. (2020) used ELM for short-term wind power output prediction and improved the training speed by more than 140 times compared to ANN methods on different time scales and seasons. Wu et al. (2020) considered the impact of different weather conditions on PV predictions and used the kernel extreme learning machine (KELM) combined with the firefly algorithm (FA) and the variational modal decomposition algorithm (VMD), resulting in a normalized root mean square error (NRMSE) and normalized mean absolute error (NMAE) below 10% for all weather conditions. Guermoui et al. (2021) improved the accuracy of ELM in multisegment ultrashort-term prediction by decomposing PV power series according to frequency range through an iterative filter (IF) decomposition method. Li et al. (2021) proposed an enhanced crow search algorithm (ENCSA) for ELM parameter optimization, which resulted in a mean absolute percentage error (MAPE) consistently below 4% in short-term wind power prediction. Another popular research direction is the combination of two machine learning models. Gastón et al. (2010) combined k-nearest neighbor (K-NN) and SVM to predict solar radiance with a substantial improvement in prediction accuracy compared to traditional methods. Shi et al. (2013) proposed a short-term wind power prediction model combining radial basis function neural network (RBFNN) and least squares support vector machine (LS-SVM), which achieved excellent results in ultra-short-term wind power generation prediction 15 min in advance. Yang et al. (2014) and Dong et al. (2015) used ANN combined with SVM for PV prediction. The results show that the prediction error of the hybrid method is smaller than that of the traditional single machine learning method. Theocharides et al. (2021) proposed a framework for a PV generation forecasting method combining ANN and K-means, which showed good performance in terms of prediction accuracy and stability. Lu et al. (2021) proposed a combined wind power prediction model based on ELM and LS-SVM, which demonstrated that the method can effectively improve prediction accuracy using historical data from a wind farm in Ningxia Hui Autonomous Region, China.

Ensemble learning models combine multiple single machine learning models to obtain better accuracy and generalization. Commonly used ensemble strategies include bagging, boosting, and stacking (Voyant et al., 2017). The bagging strategy focuses on reducing the variance by training multiple parallel base learners. Random forest is a bagging ensemble algorithm that uses decision trees as base learners, which shows good stability and generalization ability in the field of wind power prediction (Mahmud et al., 2021; Ziane et al., 2021) and can cope with random fluctuations in historical data or interference from unrelated weather factors (Lahouar et al., 2017; Shi et al., 2018). The boosting strategy focuses on reducing bias and is an additive method that continuously trains new learners. Xiong et al. (2022) and Fan et al. (2022) significantly improved the accuracy of short-term wind and PV power prediction using the XGBoost algorithm. The stacking strategy differs from bagging and boosting in that its base learners are usually heterogeneous, and it uses meta-learning to effectively combine multiple base learners for prediction. Chen and Liu (2020) and Rosa et al. (2022) improved the accuracy of wind power output prediction with a stacking ensemble strategy that takes advantage of the strengths of each prediction model in extracting different features.

Deep learning is a new research direction in the field of machine learning. Deep learning methods can automatically extract high-dimensional features from data. Previous research mostly used ANN models for power generation prediction (Izgi et al., 2012; Dumitru et al., 2016; Son et al., 2018). However, compared to ANNs with independent inputs, recurrent neural networks (RNNs) can better exploit the dependencies in time series. Pang et al. (2020) demonstrated that the RNN method improves the normalized mean bias error (NMBE) and RMSE by 47% and 26%, respectively, compared to the ANN method for short-term solar radiation prediction. Although the traditional RNN method can make good use of the information of the data, it suffers from the problems of short-term memory and gradient instability. Jung et al. (2020) demonstrated the effectiveness of LSTM-RNN for long-term prediction using an RNN model containing LSTM units for more than 63 months of data generated from multiple PV plants. Mellit et al. (2021) conducted experiments on PV power prediction over four different short-term time horizons and demonstrated that LSTM performed the best. Ahn and Park (2021) introduced LSTM units into a deep RNN for ultrashort-term and short-term prediction, and the prediction accuracy was above 92% in all cases. Li et al. (2020) introduced attention mechanisms in the traditional LSTM to reduce the effects of random variations in wind power. Luo et al. (2021) combined physical laws and domain knowledge to impose constraints on LSTM to reduce unreasonable PV power prediction results and improve the robustness of the predictions. Shahid et al. (2021) introduced a GA into LSTM and improved the accuracy of wind power prediction by 6%–35% compared to existing techniques. Hu and Chen (2018) used a combined model of LSTM and ELM to predict wind speed and outperformed SVM, LSTM, and ELM models alone in predictions 10 min and 1 h in advance. Carrera et al. (2020) used deep feedforward networks (DFN) and RNN trained on historical weather forecast data and demonstrated that the models are more accurate in predicting PV generation 1 day ahead than a single machine learning method. Hossain et al. (2021) proposed a hybrid model based on convolutional layers, gated recurrent unit (GRU) layers, and a fully connected neural network for wind power prediction, which reduced MAE, RMSE, and MAPE by 2.11%, 0.72%, and 16.88%, respectively, for wind power generation from Australian capital wind farm compared to RNN. Meng et al. (2022) combined a convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) to propose an ultra-short-term wind power prediction model and demonstrated that the model has lower prediction error and higher solving efficiency. The methods mentioned above are summarized in Table 1.

Hybrid machine learning methods have the advantage of being able to exploit the strengths of different methods to improve the prediction of machine learning models. However, intelligent optimization algorithms are often uncontrollable and slow to converge. It is also a challenge to choose the most suitable model for hybridization. The advantage of ensemble learning is the ability to organically combine multiple single machine learning models to obtain more accurate, stable, and robust results. However, the disadvantages are increased complexity and time consumption. The advantage of deep learning is its powerful learning ability, so it can handle many complex problems and a large amount of data. However, it requires powerful computational resources, and the model is difficult to build and train.

Statistical methods mainly include persistence methods, regression methods, exponential smoothing methods, and time series methods. Most research works in the field of wind power forecasting use time series methods (Das et al., 2018).

Rajagopalan and Santoso (2009) used the autoregressive moving average (ARMA) to forecast wind power within 1 hour. Gao et al. (2009) introduced an autoregressive conditional heteroskedasticity model (ARCH) based on the ARMA model to solve the problem of variance variation of wind speed time series, which has higher prediction accuracy than the classical ARMA model in wind power prediction. The limitation of ARMA is its inability to handle nonstationary series, which restricts its performance on complex forecasting tasks (Diagne et al., 2013). Abdelaziz et al. (2012) used an autoregressive integrated moving average (ARIMA) model to obtain a greater than 50.49% improvement in accuracy over the persistence model for short-term wind power prediction. Zhang et al. (2014) used the ARIMA model for short-term PV power prediction that outperformed two machine learning methods, the RBFNN, and the LS-SVM; Dokuz al. (2018) used the DBSCAN clustering algorithm to preprocess the original data before ARIMA prediction and reduced the RMSE by 20% in long-term wind speed prediction; Li et al. (2014) proposed ARMAX for PV power prediction by using relevant meteorological variables such as temperature, precipitation, and sunshine duration as exogenous inputs, and demonstrated that ARMAX can reduce the RMSE of the ARIMA and RBFNN models by 26.7% and 22.8%, respectively; Lydia et al. (2016) proposed a nonlinear ARMAX that outperformed linear ARMAX in short-term wind speed prediction; Cadenas et al. (2016) used a nonlinear autoregressive model with exogenous inputs (NARX) to predict future wind speeds, and the prediction was more accurate than a univariate method such as ARIMA. The methods mentioned above are summarized in Table 2.

In contrast to statistical methods, machine learning methods have been widely used in recent years. In terms of the time horizon of prediction, current research mostly focuses on short-term power generation prediction, in which both statistical and machine learning methods show good performance, but their performance in long-term prediction needs to be improved. In terms of modeling difficulty, statistical methods models do not require large amounts of data to build models, while machine learning models require large amounts of data to train. In terms of applicability, statistical methods describe the relationship between current and historical values. However, this also means that these methods are only suitable for predicting phenomena related to historical data, and most models can only handle linear relationships (Demolli et al., 2019). Machine learning methods are capable of handling complex multivariate prediction problems and capturing nonlinear relationships for a wider range of applications (Donadio et al., 2021; Rahman et al., 2021). Cocchi et al. (2018) demonstrated that machine learning methods have higher prediction accuracy than statistical methods in complex power generation prediction problems. Statistical methods have good interpretability, but machine learning methods are often considered to be “black box problems” (Kane et al., 2014; Alsaigh et al., 2023).

Several studies have combined the advantages of machine learning methods in fitting nonlinearities with the advantages of statistical methods in fitting linearities. Cadenas and Rivera (2010) and Jiao (2018) both developed hybrid models of ARIMA models and ANNs. ARIMA was first used to make wind speed predictions and then the obtained errors were used to build neural networks, which predicted wind speed with higher accuracy than the ARIMA and ANN models working separately. Wu and Chen (2011) combined ARMA and time delay neural network (TDNN) to predict solar radiation and also achieved good results. In addition, Liu et al. (2012) proposed the use of ARIMA to determine the structure of ANN, and the MAPE was reduced by 27.38% in multi-step prediction compared to that predicted by ARIMA alone. Zhang et al. (2020b) combined a hybrid machine learning model with an ARIMA model to predict the wind speed for the next 4 h and demonstrated that the prediction error of this method was significantly reduced.