Short-term Wind Speed Forecasting Based On Deep Learning

In recent years,due to the continuous consumption of traditional energy such as coal,oil,and natural gas,and the issues of environmental pollution and climate warming caused by them,the inexhaustible renewable energy has been widely concerned in the world.Among them,wind energy is regarded as one of the most promising energy sources because of its pollution-free,low cost and other pro-environmental characteristics.However,in the process of wind power generation,the inherent randomness and volatility of wind speed have a great impact on the conversion efficiency and stability of wind energy.To help the power department to dispatch and distribute power resources reasonably,manage and maintain the wind farm equipment regularly to reduce the economic losses caused by the imbalance of power supply and demand,equipment failure and other issues,forecasting wind speed in advance is an effective solution.With the rapid development of artificial intelligence technology,some short-term wind speed forecasting models based on deep learning have been proposed to improve the forecasting accuracy.However,in the process of wind speed forecasting,due to the significant differences in the volatility characteristics of wind speed series of different wind farms,it is difficult for individual deep learning forecasting model to maintain accurate and robust forecasting performance in multiple wind farms.Moreover,for some newly built wind farms or those with temporary failure of wind speed acquisition equipment,there is no sufficient data to train the deep learning model,and get accurate and reliable wind speed forecasting results.In view of the above problems,this paper aims to construct the short-term wind speed forecasting model based on deep learning by considering the completeness of wind farm historical data.The contributions of this study are summarized as follows:(1)A short-term wind speed forecasting model based on deep learning and ensemble learning is proposed for wind farms with complete historical data.In this study,by using the respective advantages of deep learning and ensemble learning,an ensemble deep learning model is constructed to enhance the generalization performance of the forecasting model.To improve the accuracy of the ensemble model,the two stages of ensemble learning(i.e.,the construction stage and the combination stage of basic predictor)are optimized.In the stage of basic predictor construction,three widely used deep learning models are selected as basic predictors(including multi-layer perceptron network,convolutional neural network,and long short-term memory network),and an extended AdaBoost algorithm is proposed to enhance the diversity and accuracy of the basic predictors of the ensemble model.Furthermore,a new dynamic error correction method is proposed to reduce the forecasting error of each basic predictor.In the stage of basic predictor combination,the kernel ridge regression is used as the meta-predictor to ensemble all the basic predictors.To reduce the complexity of the ensemble model,two strategies(i.e.,opposite population initialization strategy,and cross-population intercrossing strategy)are proposed to improve the solving ability of the basic non-dominated sorting genetic algorithm,and the improved algorithm is employed to conduct ensemble pruning.Finally,to verify the forecasting performance of the proposed model,two real-world wind speed datasets,four evaluation metrics,and 16 baseline models are used to establish a series of comparative experiments.Experimental results show that the proposed model can effectively improve the accuracy,stability,and generalization ability of short-term wind speed forecasting compared with the baseline models.(2)A short-term wind speed forecasting model based on deep learning and spatiotemporal correlation method is proposed for wind farms with incomplete historical data.Although the ensemble learning method can improve the forecasting performance of the deep learning model,in the case of incomplete historical data of wind farms,the training of the deep learning model will be limited,and it is difficult to obtain satisfactory forecasting results.Therefore,a spatiotemporal correlation model based on deep learning is constructed in this study,which uses multiple meteorological factors data of adjacent wind farms to train the deep learning model,and uses various spatiotemporal relationships between meteorological factors of each wind farm to depict the wind speed series features of the target wind farm.Meanwhile,to improve the spatiotemporal feature extraction ability of the deep learning model,a “multi-input,single-output” combination strategy is proposed in this study.To verify the effectiveness of the proposed model,real-world wind speed datasets from three different regions,eight evaluation metrics,and 12 baseline models are used to establish a series of comparative experiments.The experimental results show that the proposed model has better performance than other baseline models in terms of forecasting accuracy and stability.Therefore,the research on short-term wind speed forecasting based on deep learning in this study has important theoretical and practical significance for improving the energy conversion efficiency of wind farms,and solving the energy crisis.