A novel multi-task learning model based on Transformer-LSTM for wind power forecasting

Abstract

The integration of multi-task learning into multi-step deterministic and probabilistic prediction frameworks plays a pivotal role in augmenting the accuracy of wind power forecasts and mitigating associated operational uncertainties. Current wind power forecasting research usually focuses on single-task deterministic forecasting, but there is less research on multi-task learning for multi-step deterministic and probabilistic forecasting. To achieve this, a novel multi-task learning model based on Transformer-Long short-term memory (LSTM) for multi-step deterministic and probabilistic wind power forecasting. First, a novel deep learning hybrid approach (DLH) based on the dilated causal convolutional network, Transformer, LSTM, and L2 regularization is proposed to extract multi-dimensional and complex nonlinear features of wind power time series. Then, the DLH is introduced into the task-sharing layer of the multi-task learning model for multi-step deterministic prediction. In addition, a probabilistic forecasting model that integrates the proposed multi-task learning model and quantile regression is formulated to describe the forecast uncertainty of wind power. Finally, to improve the prediction accurateness, a new heuristic algorithm, namely hybrid Cauchy mutation-based Elk herd and sand cat swarm optimization, is proposed to optimize the model hyperparameters. The proposed deterministic and probabilistic forecasting models have been applied to operational datasets from a northwest China wind farm. Compared to 23 state-of-the-art deterministic models, the proposed model reduces the mean absolute error by a minimum of 0.3174 and a maximum of 9.190, with an average reduction of 2.278. Experimental outcomes also substantiate that the proposed probabilistic model exhibits superior interval sharpness when juxtaposed against six advanced benchmarks.