Hybrid CNN–SVM and Multi-Strategy Collaborative Optimization for Secondary System Configuration in Smart Grid Substations

Abstract

This paper proposes a hybrid model that integrates convolutional neural networks and support vector machines, and combines multi strategy collaborative optimization to address the complexity and dynamism of secondary system configuration tasks in smart grids. The system is based on multi-source operational data and constructs a three-stage process of "feature extraction model training configuration output". The CNN part adopts a three-layer convolution and pooling structure (convolution kernel size 3 × 3, ReLU activation) to extract topology and load features; The SVM part uses radial basis kernel functions to classify and optimize high-dimensional features. During the training process, set the learning rate to 0.001, batch size to 128, iteration times to 500, and evaluate the model's generalization performance through five-fold cross validation. The algorithm was trained using 1000 scheduling instances from 3 substations for simulation verification. The configuration accuracy reached 96.8%, which is 12.4% higher than manual experience configuration. The average response time was shortened to 0.42 seconds, and the error rate was stably controlled within 2.1%. In terms of system integration, a modular deployment structure is designed to support closed-loop operation of inference calculation, configuration generation, and result feedback. It is compatible with adaptive configuration parameters at different voltage levels such as 110kV and 220kV. In comparative testing, under consistent operating conditions, the configuration efficiency of this method increased by about 39%, and the system ran continuously for 72 hours without any configuration deviation or interruption, demonstrating good stability. Research has shown that the CNN-SVM fusion model has significant advantages in extracting features and optimizing classification, while the modular integration of various strategy optimization architectures and systems has the effect of improving setup efficiency and trustworthiness. This study integrates CNN-SVM, GA/PSO, reinforcement learning, and graph neural networks to form a comprehensive strategy optimization system suitable for the secondary system setting of substations. Unlike previous separate applications of CNN or SVM, this study highlights the synergistic effect under complex constraints and emphasizes the online regulation effect and multi-level voltage promotion capability. Moreover, compared to existing AI optimization applications in other fields, this article focuses more on engineering implementation and real-time constraints in power scenarios, thus differentiating it from existing methods.