MLP sliding-window forecasting for electricity load prediction: a multi-scale evaluation using an ONNX-based Java framework

Deep learning models for time-series forecasting offer significant potential but face deployment challenges. This paper advances a modular framework for deploying ONNX models in Java, building on previous work. The research paper assesses the MLP sliding-window architecture through different time intervals which extends the previous LSTM-based system to establish primary performance standards that future research can use for comparison. The framework incorporates additional features for enhanced evaluation, including five metrics (MAE, RMSE, MAPE, SMAPE, R²), automated archiving, and high-quality graphical exports. Evaluations using electricity consumption data (ETTh1) and benchmark datasets show that the MLP sliding-window model outperforms LSTM in terms of R², MAE, and MAPE, achieving scores of 0.9895, 0.5907, and 7.19%, respectively, indicating high accuracy across various domain. The framework also implements a custom threshold (ε = 1e−10) to handle near-zero values in MAPE calculation, ensuring reliable result storage for reproducibility. The findings reveal that simpler MLP designs can either match or exceed LSTM performance in specific scenarios while offering faster processing and easier implementation. Detailed comparisons and guidelines for deployment are included, along with insights into model selection criteria.