Solar irradiation time series forecasting for Wireless Sensor Networks via hybrid BiLSTM–1D CNN and feature-enriched inputs

Energy sustainability in Wireless Sensor Networks (WSNs) remains a critical challenge, particularly in remote or inaccessible environments where battery replacement is impractical. Solar energy harvesting offers a promising and renewable alternative; however, its inherent intermittency and variability require accurate forecasting to enable efficient energy management and system reliability. In this work, we propose a comprehensive solar irradiance forecasting framework based on 20 years of hourly data, integrating advanced feature engineering and hybrid predictive modeling. To enhance signal quality and capture multi-scale temporal characteristics, a set of complementary time-series smoothing techniques — Simple Moving Average (SMA), Exponential Moving Average (EMA), Weighted Moving Average (WMA), and Hull Moving Average (HMA) — is employed to preprocess the raw data. The core contribution is a hybrid BiLSTM–1D CNN–XGBoost architecture, designed to jointly exploit long-term temporal dependencies, short-term local patterns, and nonlinear residual relationships. Specifically, the 1D-CNN extracts local temporal features, the BiLSTM captures bidirectional long-range dependencies, and XGBoost operates on the learned feature space to refine predictions through gradient-boosted regression. The proposed model is extensively evaluated against several baseline approaches, including ARIMA, 1D-CNN, LSTM, BiLSTM, and GRU. Experimental results demonstrate that the hybrid framework significantly improves forecasting accuracy and robustness across multiple evaluation metrics. This study contributes to the literature by introducing a feature-enriched hybrid forecasting strategy tailored for solar-powered WSNs, providing a practical and scalable solution for proactive energy-aware management in IoT and edge-based systems.