Scholarlink

Accurate short-term forecasting of household energy consumption is a fundamental premise for efficient smart-grid operation and sustainable planning. Many conventional statistical models, such as ARIMA and regression, fail to learn nonlinear, long-term temporal dependencies intrinsic in energy data. In this work, we propose a multivariate time-series forecasting approach that incorporates Recursive Feature Elimination coupled with an LSTM network implemented in Keras. The architecture first selects the most informative inputs (e.g., voltage, reactive power, and sub-metering) and subsequently learns temporal dynamics in order to forecast short-term demand. Using the UCI Individual Household Electric Power Consumption dataset, we compare the proposed model against a strong Random Forest baseline. In its present configuration, the Random Forest yields lower MAE and RMSE on the test set; nevertheless, the RFE–LSTM still achieves a non-trivial coefficient of determination ( \(R^2 \approx 0.60\) ) and enjoys feature selection offering a transparent and extensible framework for smart-grid energy management. The Adam optimizer with a learning rate of \(1\times 10^{-3}\) , \(\beta _1=0.9\) , and \(\beta _2=0.999\) , which minimized the mean squared error loss with early stopping on validation loss, was used for training the model.

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

Multivariate Time-Series Prediction of Energy Consumption Using Recursive Feature Elimination (RFE) and LSTM–Keras

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Prischa Bajeli,
  • Sidharth Praveen,
  • Swara Kadam,
  • Shilpa Sonawani

摘要

Accurate short-term forecasting of household energy consumption is a fundamental premise for efficient smart-grid operation and sustainable planning. Many conventional statistical models, such as ARIMA and regression, fail to learn nonlinear, long-term temporal dependencies intrinsic in energy data. In this work, we propose a multivariate time-series forecasting approach that incorporates Recursive Feature Elimination coupled with an LSTM network implemented in Keras. The architecture first selects the most informative inputs (e.g., voltage, reactive power, and sub-metering) and subsequently learns temporal dynamics in order to forecast short-term demand. Using the UCI Individual Household Electric Power Consumption dataset, we compare the proposed model against a strong Random Forest baseline. In its present configuration, the Random Forest yields lower MAE and RMSE on the test set; nevertheless, the RFE–LSTM still achieves a non-trivial coefficient of determination ( \(R^2 \approx 0.60\) ) and enjoys feature selection offering a transparent and extensible framework for smart-grid energy management. The Adam optimizer with a learning rate of \(1\times 10^{-3}\) , \(\beta _1=0.9\) , and \(\beta _2=0.999\) , which minimized the mean squared error loss with early stopping on validation loss, was used for training the model.