Fusion of LSTM and SVM to optimize data anomaly detection and demand forecasting in power marketing business expansion

Abstract

This paper proposes a collaborative optimization framework that integrates LSTM (Long Short-Term Memory) and SVM (Support Vector Machine) to solve the problem of load data anomaly detection and demand forecasting in power marketing business expansion. The robustness of time series features is enhanced by three-layer stacked LSTM, combined with dual classifier SVM dynamic boundary optimization, and the attention mechanism is introduced to realize the adaptive weighting of key features. An alternating iterative training strategy and a sliding window real-time update mechanism are designed to improve the model's responsiveness to load fluctuations and abnormal patterns. Experiments show that this method outperforms benchmark models such as isolation forest and GRU (Graphics Processing Unit) in anomaly detection accuracy (90.4%) and prediction MAPE (Mean Absolute Percentage Error) (7-day cycle error), and has real-time detection feasibility (GPU latency 1.1ms/sample). The research provides new ideas for dynamic feature fusion and decision boundary optimization in power marketing scenarios.