Real-Time Web Inferencing of a Bilstm-Informer Hybrid Model with Autoregressive Features Optimization for Improved Photovoltaic Power Output Forecasting

Abstract

To ensure an efficient photovoltaic (PV) renewable energy grid, it is essential to address the uncertainty inherent in the power systems. Accurate real-time forecasting is critical for enabling smarter energy management, grid stability, and effective resource scheduling. This study proposes a BiLSTM-Informer hybrid model that significantly improves multi-step PV power output prediction. Using a 39.2 kWp PV system and location-specific weather data, the model’s performance is benchmarked against modelled machine and deep learning forecasting techniques such as Lasso Regression (R²:0.611), K-Nearest Neighbours (0.625), Support Vector Regression (0.680), and Extreme Gradient Boosting (0.810), which underperformed due to limited capacity to capture temporal dependencies. Ensemble learning improved short-term accuracy (R² = 0.861) but lacked adaptability for long-term patterns. Deep learning models like GRU (R²:0.874) and LSTM (R²: 0.899) improved sequential learning but could not dynamically prioritize feature relevance. The proposed BiLSTM-Informer model combines BiLSTM’s ability to handle bidirectional temporal sequences with the Informer’s sparse attention mechanism for long-range forecasting. Advanced feature engineering, including Fourier transformation, cyclic encoding, and autoregressive lag features optimization, further enhanced performance. The model achieved a mean absolute error of 1.22 kWh, RMSE of 2.21 kWh, and R² of 0.952. This reflects a 20.1% increase in forecasting accuracy and a 30% reduction in computational redundancy. For practical validation, the model was deployed using a Streamlit web interface on Orender. Real-time forecasts demonstrated high accuracy (89–97.3%) with low latency, confirming its robustness for deployment and operational use. This work sets a benchmark for intelligent PV forecasting and grid management.

Keywords: Cyclic Encoding, Energy Management System, Features Optimization, Photovoltaic Power Forecasting, Real-Time Inferencing