An Integrative Hybrid Feature Fusion Framework Using XGBoost–BiLSTM for Enhanced Solar Power Forecasting Accuracy

Abstract

This paper proposes an Integrative Hybrid Feature Fusion Framework that integrates the temporal sequence learning and the structural data processing capacity of Bidirectional Long Short-Term Memory (BiLSTM) network and eXtreme Gradient Boosting (XGBoost), respectively. The proposed approach begins with 24 hours sliding window strategy and robust data pretreatment of solar energy terms including ambient temperature, wind speed, and Global Horizontal Irradiance (GHI). As a result of the architectures, the preprocessed data is split into two parallel blocks. The XGBoost branch implements dense projection, leaf index embedding, and tree-based modeling to produce a 64-dimensional feature vector that can capture intricate feature interactions, even in their non-linear form. During this process, the deep temporal relationship between sequential data is extracted by the BiLSTM branch and then a 64-dimensional context vector is output. The representations are then combined into a 128-dimensional representation and input to a fully connected fusion prediction head, yielding the final normalized regression output. The inverse transformation is then applied to produce the system prediction of the final Direct Current (DC) power (kW) which is fully assessed with the use of a number of standard error measures. Inspired by the complementary strengths of both approaches, the framework presents a hybrid method that yields a significant improvement on the forecast accuracy of solar power. Taking advantage of the complementary attributes of both approaches, the structure is adapted into a hybrid method which effectively merges static structural abstractions with dynamic temporal conditions, resulting in improved forecast accuracy of solar power. All the suggested models had excellent predicting accuracy with the Mean Absolute Error (MAE) of 0.912 KW, Mean Squared Error (MSE) of 2.579 KW, Root Mean Squared Error (RMSE) of 1.643 KW, and Mean Absolute Percentage Error (MAPE) of 1.948%.