Application of parallel computing in forecasting problems of complex engineering processes

Abstract

Complex engineering processes require short-term prediction to maintain the stability of energy-intensive industrial operations under severe computational and latency constraints. The research is a 24-hour-ahead multi-step prediction of aggregated electrical load in one of the Ukrainian metallurgical plants, where high-frequency SCADA data are sampled at 15-minute intervals, totaling 105,120 observations over 2022-2025. This is to predict energy demand with a maximum inference latency of 250 ms to enable real-time dispatch coordination. It is an advanced model that uses mixed-precision (FP16) computing and asynchronous data loading on GPU-accelerated data parallelism (CUDA 12.2, PyTorch 2.1, NVIDIA RTX A6000 48 GB). A leakage-conscious walk-forward validation protocol performs five repeated validations and guarantees temporal consistency and reproducibility. Forecast performance is evaluated using mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and R². At the same time, computational efficiency is assessed in terms of speedup, throughput, and inference latency. The proposed parallel hybrid framework reduces the MAE to 2.41 MW, the RMSE to 3.05 MW, the MAPE to 2.97%, and the R² to 0.98, outperforming the ARIMA, standalone LSTM, and Transformer baselines. GPU acceleration achieves a 6.67-fold reduction in training time, increases throughput from 415 to 1964 samples per second, and lowers inference latency from 312 ms to 87 ms, satisfying operational constraints. Parallel deep forecasting can support real-time industrial decisions, improve load balancing, and enhance energy resilience in Ukraine.