A novel GBDT-BiLSTM hybrid model on improving day-ahead photovoltaic prediction

Abstract

Despite being a clean and renewable energy source, photovoltaic (PV) power generation faces severe challenges in operation due to its strong intermittency and volatility compared to the traditional fossil fuel power generation. Accurate predictions are therefore crucial for PV’s grid connections and the system security. The existing methods often rely heavily on weather forecasts, the accuracy of which is hard to be guaranteed. This paper proposes a novel GBDT-BiLSTM day-ahead PV forecasting model, which leverages the Teacher Forcing mechanism to combine the strong time-series processing capabilities of BiLSTM with an enhanced GBDT model. Given the uncertainty and volatility inherent in solar energy and weather conditions, the gradient boosting method is employed to update the weak learner, while a decision tree is incorporated to update the strong learner. Additionally, to explore the correlation between photovoltaic power output and historical time-series data, the adaptive gradient descent-based Adam algorithm is utilized to train the bidirectional LSTM model, enhancing the accuracy and stability of mid- to long-term time-series predictions. A prediction experiment, conducting with the real data from a PV power station in Sichuan Province, China, was compared with other methods to verify the model’s effectiveness and robustness.