Research and Optimization of Charging Pile Demand Forecasting Model Based on Data Mining

Abstract

The rapid growth of electric vehicles (EVs) has led to a significant increase in the demand for accurate forecasting of charging pile usage, essential for optimizing infrastructure planning and energy management. Effective prediction models help to balance grid loads, reduce waiting times, and enhance the overall charging experience. Existing methods, based on traditional Machine Learning (ML) or basic neural network models, struggle to capture complex spatial-temporal relationships and multivariate dependencies. To address these limitations, this research proposes an advanced hybrid Mandrill-tuned Convolutional Long Short Memory with Auto Encoder (MCLSM-AE) framework that combines the Mandrill Optimization Algorithm (MOA), Auto Encoder (AE), and CNN-LSTM with an Attention Mechanism. The novelty of this approach lies in integrating MOA for optimal hyper parameter tuning, AE for dimensionality reduction and CNN-LSTM for spatial-temporal demand modeling, enhanced with an attention mechanism for improved interpretability. The model is trained on a dataset comprising historical EV charging data, traffic patterns, weather information, and spatial grid mappings. Preprocessing steps include data normalization and Missing Value Imputation (MVI) to ensure data quality. The proposed model workflow involves reducing data dimensionality with AE, extracting spatial patterns with CNN, and capturing temporal dependencies using LSTM, with MOA optimizing model parameters. Experimental results demonstrate the suggested MCLSM-AE model superior performance, achieving a MSE (0.00000000022) RMSE of (0.000014832), and MAE (0.0275) compared to existing methods. The research provides a robust and scalable solution for EV charging demand forecasting, addressing existing limitations and contributing to better infrastructure and energy management strategies