Parallel TimesNet-BiLSTM model for ultra-short-term photovoltaic power forecasting using STL decomposition and auto-tuning

Accurate forecasting of photovoltaic (PV) generation can mitigate the impact of weather stochasticity on power systems and facilitate the development of effective energy scheduling strategies. This study presents a parallel TimesNet-bidirectional long short-term memory (PA-TimesNet-BiLSTM) model for ultra-short-term PV power forecasting. Initially, a seasonal trend decomposition using loess (STL) method was employed to decompose the raw data and reconstruct the input features. Subsequently, the TimesNet model extracts multiple periodic features in a two-dimensional space, while the BiLSTM model addresses long-term data dependencies. The PA-TimesNet-BiLSTM model hyperparameters are optimized using an asynchronous successive halving algorithm. The evaluation utilized standard metrics and Diebold–Mariano testing to assess the predictive performance of 12 benchmark models across two datasets. The results demonstrate the competitive performance of the PA-TimesNet-BiLSTM model. STL decomposition significantly benefits PV power forecasting. On the Australian dataset, the mean absolute error (MAE) and root mean square error (RMSE) of the PA-TimesNet-BiLSTM model improved by 7.42% and 4.31%, respectively, compared to the serial TimesNet-BiLSTM model. The STL-PA-TimesNet-BiLSTM achieved reductions of 29.05% and 33.01% in MAE and RMSE, respectively, compared with the PA-TimesNet-BiLSTM model. The PA-TimesNet-BiLSTM model effectively captured multidimensional periodic data features, enhancing its applicability to diverse prediction tasks.