A novel decision support system for enhancing long-term forecast accuracy in virtual power plants using bidirectional long short-term memory networks

Accurate forecasting of power generation is a serious challenge of virtual power plant (VPP) in day ahead (DA) market because of the volatility and uncertainty of renewables. The recursive prediction technique used in bidirectional long short-term memory (BiLSTM) network often struggles with long-term accuracy. This study proposes a novel decision support system (DSS) to generate unknown future inputs, called “DSS test data”, in the recursive prediction technique and tackle the long-term forecasts limitation. The proposed DSS integrates the K-means clustering algorithm and the least squared optimization method. The K-means clustering algorithm classifies historical data into five distinct day types—rainy, overcast, partly cloudy, cloudy, and sunny—based on maximum daily power generation. The DSS employs least squared optimization method to refine the DSS test data for the BiLSTM model, utilizing the most recent seven days of data. Additionally, this study incorporates a variable lookback period within the BiLSTM model to enhance the accuracy of the forecasting model. The DSS-BiLSTM model forecasts VPP power generation 38 h ahead in the Japanese DA power market. Compared to BiLSTM, LSTM, transformer network, attention-based network, gated recurrent unit, and five statistical time series models, the proposed model demonstrates superior accuracy and reduced dispersion in long-term forecasts. The daily mean absolute error for the DSS-BiLSTM, BiLSTM, LSTM, transformer network, attention-based network, and gated recurrent unit models, for a 38-h forecast horizon, are 0.26 GW, 0.48 GW, 0.45 GW, 0.69 GW, 0.66 GW, and 0.62 GW, respectively. This pattern is consistent across the three other error metrics and various forecasting time horizons, indicating that the DSS-BiLSTM model consistently outperforms the other models evaluated in this study in terms of prediction accuracy. The main advantages of the proposed model include ease of implementation, low dispersion, and high forecasting accuracy across various settlement periods, as evidenced by multiple accuracy metrics.