Incrementally trained short-term wind turbine power prediction model based on long short-term memory

Abstract

To enhance the prediction accuracy of short-term wind turbine power prediction models, this study proposes an incremental prediction model based on Long Short-Term Memory (LSTM). Initially, the LSTM module is employed to process data related to wind turbines, and the key parameters of the network (hidden layer units, learning rate, batch size, time step) are estimated using the discrete particle swarm optimization algorithm (DPSO) Subsequently, incremental learning is introduced to dynamically update the model, and a merged weight updating strategy is adopted to alleviate potential overfitting during the incremental training process. In this study, publicly available wind energy datasets are used for experimentation (with a data time interval of 5 minutes), and compared and validated against three other framework models: overall LSTM, Informer, and Paddle. Experiments show that the MAE value of the proposed model is 0.137 and the RMSE value is 0.199, which is comparable to the performance of the overall LSTM model (with fluctuations of 2% and 0.7% respectively). The average training time for the proposed model is 14,773.4 ms, representing an 81% reduction compared to other models.