Yibin Zhao
View Full Paper
Temporal Convolutional Network, Deep Belief Network, Power System, Short-Term Load Probability Forecasting, Risk Quantification
In this study, we propose a hybrid TCN-DBN architecture for proba- bilistic short-term load forecasting (STLPF) and dynamic risk quan- tification in power systems. Traditional deterministic forecasting methods fail to capture load uncertainty and provide actionable risk insights. By integrating the temporal dependency modeling capabil- ity of temporal convolutional networks (TCN) with the deep prob- abilistic representation of deep belief networks (DBN), our frame- work achieves joint optimization of forecasting accuracy and uncer- tainty quantification. The TCN module employs causal and dilated convolutions to extract long-range spatio-temporal features, while the DBN module generates probabilistic load distributions through Bayesian inference. Experimental results on the EUNITE dataset demonstrate that the proposed model reduces MAPE by 89.7% com- pared to standalone TCN and achieves a 90% prediction interval cov- erage probability (PICP) with a normalized interval width (PINAW) of 3.04%, outperforming state-of-the-art models (e.g., ResNet-LSTM, TCN-Transformer). The framework provides critical support for grid resilience planning and risk-informed decision-making under uncer- tain load conditions.
Important Links:
Go Back
