Design and Implementation of Short-Term Load Forecasting using STM

Residential energy consumption constitutes a significant portion of the total electricity demand, emphasizing the urgent need for accurate short-term load forecasting to facilitate efficient energy management. Initially, a Bidirectional Long Short-Term Memory (BiLSTM) model was employed to analyze household energy consumption patterns using publicly available datasets. To further enhance forecasting accuracy and improve adaptability to real-world scenarios, a real-time data collection system was developed utilizing an ESP32. This system was designed to capture key parameters, including voltage, current, power, and energy consumption, thereby generating a custom dataset. Subsequently, this custom dataset was used to train the BiLSTM model, which was then deployed on an STM32 microcontroller for edge-based forecasting. For fine-tuning the model's hyperparameters, the Particle Swarm Optimization (PSO) technique was implemented. Pre-processing techniques were applied to filter and reduce noise within the datasets. Comparative studies conducted between the BiLSTM models trained on publicly available data and the customized data demonstrated the superiority of the customized data in terms of forecasting accuracy and sensitivity in edge-based performance. The proposed methodology outperforms traditional forecasting techniques by enabling a scalable, adaptive, and effective solution for residential energy management.