Load Forecasting And Modeling For Power System

Accurate load forecasting and modeling play a pivotal role in ensuring the stability, reliability, and economic efficiency of modern power systems. With the increasing integration of renewable energy sources, distributed energy resources, and demand-side management strategies, power systems are becoming more dynamic and complex, making traditional load forecasting methods inadequate. This dissertation introduces two novel approaches to address the challenges associated with day-ahead load forecasting and load modeling. First, a Diffusion Model-Based Probabilistic Day-Ahead Load Forecasting (PDALF) Framework is proposed to enhance the accuracy and robustness of load forecasting. By employing a conditional denoising diffusion probabilistic model (DDPM), the framework, termed DALNet (Diffusion-Augmented Load Network), generates load curves by progressively adding and removing Gaussian noise in a Markov chain. This approach effectively models the complex distribution of load data, avoiding the error accumulation inherent in rolling forecasts and capturing intra-day correlations. Additionally, a Temporal Multi-Scale Attention Block (TMSAB) is integrated into DALNet to extract both positional and temporal information, further improving prediction accuracy. Comparative experiments on real-world datasets, including GEFCom2014 and Arizona State University's load data, demonstrate that DALNet significantly outperforms traditional benchmarks such as LSTM, Transformer, and Bayesian Neural Networks (BNNs), offering superior reliability, sharpness, and overall forecasting performance. Second, this dissertation presents a Reinforcement Learning-Based Symbolic Regression (SR) Framework for load modeling to address the limitations of fixed-form parametric models and complex machine learning models that lack interpretability. Leveraging the Actor-Critic reinforcement learning architecture, a trainable expression tree structure is designed to discover mathematical expressions that describe the relationship between load characteristics and system variables. To balance model complexity and interpretability, a candidate pool is employed to refine the best-performing expressions using policy gradient optimization and gradient-based fine-tuning. Case studies demonstrate that the proposed symbolic regression framework effectively captures the nonlinear dynamics of load responses under various grid disturbances, outperforming conventional parametric models and artificial neural networks in terms of accuracy, interpretability, and computational efficiency. The combination of these two innovative approaches provides a comprehensive solution to the challenges of probabilistic load forecasting and dynamic load modeling in modern power systems. By bridging the gap between accuracy, interpretability, and scalability, this work contributes to advancing the state-of-the-art in power system analysis and operation.