A hybrid long-term photovoltaic power prediction model integrating a BiLSTM network with residual correction via CatBoost

• Long-term forecasting of PV power using a hybrid BiLSTM-CatBoost model. • BiLSTM captures temporal dependencies in DC power. • CatBoost corrects residual errors. • Validates the model in multi-weather condition. • Comparative with other models confirmed its reliability and accuracy. To improve the power output of PV solar panels and facilitate their integration into the power grid, accurate PV power forecasting is important for managing the instability of their output, given the variation in weather conditions. This study uses a hybrid approach that applies bidirectional long short-term memory (BiLSTM) with categorical boosting algorithm (CatBoost) for long-term PV power forecasting. The proposed framework aims to improve the accuracy of photovoltaic production forecasts by using BiLSTM to perform temporal forecasting and the CatBoost model for the residual correction. The BiLSTM–CatBoost hybrid model is evaluated under three real-weather conditions (sunny, cloudy, and rainy) using real-world photovoltaic data, ensuring that the assessment reflects actual operating scenarios. To validate the performance of the proposed model, we compare the BiLSTM–CatBoost hybrid against several baseline and reference architectures, including the persistence model, LSTM, BiLSTM, LSTM–XGBoost, LSTM–CatBoost, and BiLSTM–XGBoost. Using real-world PV data, the experimental results show that the BiLSTM–CatBoost model consistently achieves superior accuracy across all weather conditions, reaching a mean absolute error (MAE) of 2.808 kW and an average R2R^2R2 score of 0.9998. These comparative results demonstrate not only the robustness of the hybrid approach under diverse atmospheric scenarios but also its strong potential for reliable, real-time photovoltaic forecasting.