A CNN-BiLSTM short-term wind power forecasting model incorporating adaptive boosting

Abstract

Abstract Aiming at the wind power signal with the characteristics of intermittency, nonlinearity, volatility, non-stationarity and uncertainty, this paper establishes a wind power prediction model based on the combination of variational modal decomposition (VMD), convolutional neural network (CNN), bi-directional long and short-term memory network (BILSTM) and adaptive boosting mechanism (AdaBoost). In terms of data processing, the core parameters of VMD such as decomposition modulus number and penalty factor affect the data decomposition ability, thus the core parameters of VMD are optimized using the multi-strategy mutation sand cat swarm optimization (SSCSO). The global search ability and convergence speed of SSCSO algorithm are enhanced by integrating cubic mapping, spiral search strategy and sparrow alert mechanism, etc., and are applied to optimize the core parameters of VMD, so as to effectively improve the data decomposition performance of VMD; in terms of the prediction model, for the existence of a single deep neural network model with slow arithmetic speed, artificial parameter tuning, etc., which affects the overall prediction accuracy of the model, thus CNN-BiLSTM combination prediction model with the introduction of AdaBoost is adopted. The CNN-BiLSTM is repeatedly trained as a weak predictor and outputs the prediction results, and the weights are calculated and the errors are corrected according to the prediction error values of each weak predictor. Finally, the strong predictor is obtained by combining several groups of weak predictors after several rounds of training, and the output predicted values are superimposed to obtain the final predicted values, which further improves the overall prediction accuracy of the model, and the strong predictor composed of the CNN-BiLSTM model trained in several rounds is able to process the data more adaptively, and improves the operation speed to a certain extent under the premise of guaranteeing the prediction accuracy. The experimental results show that the root mean square error (RMSE), mean absolute error (MAE) The experimental results show that the RMSE, MAE, correlation coefficient and running time of the proposed model are better than those of SSCSO-VMD-CNN-BiLSTM, SSCSO-VMD-CNN-LSTM and SSCSO-VMD-CNN-GRU, VMD-CNN-BiLSTM-Adaboost, SABO-VMD-CNN-BiLSTM-Adaboost, DBO-VMD-CNN-BiLSTM-Adaboost and WOA-VMD-CNN-BiLSTM-Adaboost prediction models. Therefore, the combined model proposed in this paper has better prediction accuracy and running speed.