Peak-valley resource demand forecasting for enhanced data reliability

Accurate load demand forecasting serves as the fundamental prerequisite for data reliability analysis in multi-user heterogeneous load access scenarios. In such environments, the volume of data flows generated by industrial, residential, and commercial loads exhibits complex temporal variation patterns characterized by distinct periodicity and volatility. These patterns are governed by multiple factors including diverse user behaviors, operational schedules, and seasonal variations. However, conventional reliability assessment methods rarely account for the periodic nature of peak/valley load fluctuations and the volatility of sudden data surges, leading to inaccurate reliability evaluations, e.g., underestimated transmission risks during peak load periods or inefficient resource redundancy during off-peak intervals. To address this challenge, we propose a peak-valley load demand forecasting (TAPV-RDF) model. The core motivation is to model and fit the complex temporal variation patterns of heterogeneous load data flows, which demonstrate multi-scale periodicity mixed with surge components. Given the high complexity of these patterns, the proposed model first constructs a “user-load” characteristic mapping framework by integrating multi-dimensional user attributes with time-series load data. A Gated User-aware Temporal Fusion Variational Autoencoder (GUTF-VAE) is designed to extract periodic load characteristics and surge amplitude features; this component introduces gated fusion mechanisms and attention-driven latent space regularization. Subsequently, a Peak-Valley Multi-Scale Gated LSTM (PVMS-G-LSTM) is proposed to capture multi-scale temporal dependencies across varying load conditions. Experiments on three real-world datasets demonstrate that TAPV-RDF consistently outperforms state-of-the-art baselines. Specifically, our framework achieves improvements across the three datasets. Ablation studies confirm the contribution of each architectural component, while statistical significance tests validate the reliability of improvements. The proposed framework provides a practical solution for load demand forecasting and resource provisioning in modern power grid systems.