OPTIMIZATION OF HYBRID TCN, LSTM, AND LIGHTGBM MODELS FOR ENERGY CONSUMPTION FORECASTING IN SMART HOMES

Abstract

Accurate short-term load forecasting is a key task for effective energy resource management in smart home systems. Hybrid models that combine deep learning (DL) architectures and decision tree ensembles are a leading direction in modern research. An analysis of recent publications confirms that comparing Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN) networks is a popular topic, and hybridization with LightGBM and the use of error correction strategies ("residual forecasting") are proven practices for improving accuracy. However, a literature review reveals several unresolved parts of the general problem: 1) the lack of a systematic analysis of the trade-off between forecast accuracy and computational cost (training time, resource requirements), which is critical for implementation on Internet of Things (IoT) devices; 2) insufficient research on the impact of feature engineering, particularly feature selection, on the computational efficiency of hybrid models; 3) a tendency to focus on accuracy metrics without providing practical methodologies for selecting the optimal model depending on the specific task. This work aims to fill these gaps. A multi-stage experimental analysis was implemented. Two hybridization strategies were tested for the selected models: "peak correction" and "residual forecasting." A methodology for optimizing training time by selecting the most important features for the corrector model was developed. To ensure the statistical significance of the results, time-series cross-validation was applied. The study confirmed that hybrid models significantly outperform base models, and the "residual forecasting" strategy is the most effective. Two high-performance specialized configurations were identified. The LSTM + LightGBM hybrid demonstrated the highest overall accuracy (MAPE 13.36%). At the same time, the TCN + LightGBM hybrid proved to be more effective in forecasting critical peak loads (Peak Magnitude MAPE 16.71%) and was 21% faster to train. A key result is the proposed methodology for optimizing the TCN + LightGBM model through feature selection, which allowed for a 5.4-fold speed-up in training while maintaining high peak forecasting accuracy (Peak MAPE 16.77%). The work fills the identified gaps in the literature by providing not only quantitative results but also a practical methodology for the informed selection of a forecasting architecture depending on the priority tasks of a smart home system: maximum overall accuracy, priority management of peak loads, or balanced performance for resource-constrained devices. The proposed optimized hybrid approach is promising for practical implementation in adaptive energy management systems due to its proven balance of high accuracy and low computational costs.