Real-time web inference of a BiLSTM-informer hybrid model for enhanced photovoltaic power output forecasting

To ensure an efficient Photovoltaic (PV) renewable energy grid, it is essential to address the uncertainty inherent in power systems. An efficient energy management system must be capable of prioritising energy distribution based on an applicable and effective real-time forecasting of the generation output of the PV system. This study proposes a novel BiLSTM-Informer hybrid model that outperforms benchmarked machine and deep learning approaches in forecasting multi-step PV output by addressing their inability to capture non-linear temporal dependencies and lack of dynamic features weighting. A 39.2 kWp PV system serves as a case study, incorporating location-specific weather parameters. The proposed model integrates Fourier transformation, cyclic encoding, and autoregressive feature optimization to enhance pattern recognition and short-term variability. It achieved superior accuracy, with a mean absolute error (MAE) of 1.22 kWh, a root means square error (RMSE) of 2.21 kWh, and a coefficient of determination (R2) of 0.952. This reflects a 20.1 % increase in online forecasting accuracy. Unlike previous studies, this work integrates real-time web inferencing using a Streamlit interface on Orender, thereby validating the model’s robustness under live deployment. The model demonstrated forecasting accuracy ranging from 89 % to 97.3 % across multiple forecasting (1-hour to monthly) horizons with reduced computational overhead. These results position the BiLSTM-Informer as a novel benchmark for real-time PV forecasting and intelligent power grid management. The data and pre-trained models are available at the dedicated GitHub repository: https://github.com/kamilkenny/EDA and the Inferenced Model link is: https://kamil-deployment-of-edgehill-durning.onrender.com/.