Scholarlink

<p>Accurate short-term load forecasting (STLF) is crucial for reliable and efficient power system operation. This study investigates how Particle Swarm Optimization (PSO) parameter configurations affect the optimization behavior and forecasting accuracy of a hybrid Convolutional Neural Network-Multi-Layer Perceptron (CNN-MLP) model. Two real-world half-hourly electricity load datasets from New South Wales (NSW) and Queensland (QLD), Australia, are used to examine robustness under different optimization settings. PSO is used to tune key CNN-MLP hyperparameters, including the number of Conv1D filters, the number of dense-layer neurons, the batch size, and the number of training epochs. Beyond identifying a single best hyperparameter set, we analyze the influence of PSO settings, swarm size (swarmsize), maximum iterations (maxiter), and evaluation budget (number of model evaluations) on convergence behavior, computational cost, and achievable accuracy. Cross-validation results show that PSO configuration can materially change convergence efficiency and performance, reaching best-case cross-validated MAPE values of 0.0088 (0.88%) for NSW and 0.0101 (1.01%) for QLD. This sensitivity analysis provides practical guidance for reproducible PSO configuration. The findings highlight the need to carefully select PSO parameters when applying metaheuristic optimization to hybrid deep learning models for STLF. Limitations include two regional datasets and one architecture; future work will consider broader datasets and adaptive optimization strategies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating the influence of particle swarm optimization parameters on hybrid convolutional neural network–multilayer perceptron models for short-term load forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Tuan Anh Nguyen,
  • Thanh Ngoc Tran

摘要

Accurate short-term load forecasting (STLF) is crucial for reliable and efficient power system operation. This study investigates how Particle Swarm Optimization (PSO) parameter configurations affect the optimization behavior and forecasting accuracy of a hybrid Convolutional Neural Network-Multi-Layer Perceptron (CNN-MLP) model. Two real-world half-hourly electricity load datasets from New South Wales (NSW) and Queensland (QLD), Australia, are used to examine robustness under different optimization settings. PSO is used to tune key CNN-MLP hyperparameters, including the number of Conv1D filters, the number of dense-layer neurons, the batch size, and the number of training epochs. Beyond identifying a single best hyperparameter set, we analyze the influence of PSO settings, swarm size (swarmsize), maximum iterations (maxiter), and evaluation budget (number of model evaluations) on convergence behavior, computational cost, and achievable accuracy. Cross-validation results show that PSO configuration can materially change convergence efficiency and performance, reaching best-case cross-validated MAPE values of 0.0088 (0.88%) for NSW and 0.0101 (1.01%) for QLD. This sensitivity analysis provides practical guidance for reproducible PSO configuration. The findings highlight the need to carefully select PSO parameters when applying metaheuristic optimization to hybrid deep learning models for STLF. Limitations include two regional datasets and one architecture; future work will consider broader datasets and adaptive optimization strategies.