Optimizing Accuracy and Reliability in Power Load Forecasting Using a Data-Driven RF and LSTM Hybrid Framework

Effective power load forecasting(PLF) is a fundamental necessity for sustaining the operational quality and economic efficiency of modern power supply infrastructure, especially amid rising global electricity demand and the transition to smart grid systems. Traditional forecasting techniques often struggle with capturing non-linear and temporal dependencies inherent in real-world power consumption data, which are affected by diverse and dynamic factors such as climate, socioeconomic activity, and regional consumption patterns. We propose a hybrid forecasting framework that leverages the complementary strengths of Random Forest (RF) and Long ShortTerm Memory (LSTM) to achieve high-accuracy power load predictions. The RF is first employed to perform advanced feature engineering and data augmentation, enhancing the temporal and non-linear relationships in the input space. These enriched features are then fed into the LSTM model, which effectively learn long-range temporal dependencies to deliver highly accurate forecasts. Experimental evaluations conducted on real-world zone-based power consumption datasets demonstrate that the proposed data driven hybrid RF and LSTM significantly outperforms standalone RF and LSTM models across multiple performance metrics. In particular, the proposed hybrid RFLSTM model resulted in a Mean Absolute Percentage Error (MAPE) of $2.29 \%$, substantially outperforming the standalone RF ($\mathbf{1 0. 4 5 \%}$) and LSTM (9.35%) models. Likewise, in terms of the $\mathbf{R}^{\mathbf{2}}$ metric, the hybrid approach attained an accuracy of $98.13 \%$, compared to $\mathbf{5 7. 5 4 \%}$ for $\mathbf{R F}$ and $\mathbf{6 7. 4 3 \%}$ for LSTM. The findings highlight a substantial increase in the accuracy and robustness of power consumption prediction, attributable to the proposed hybrid methodology.