Multivariate-aided Power-consumption Prediction Based on LSTM-Kalman Filter

Abstract

Forecasting the power consumption of home appli-ances on a time-series basis is significant in monitoring and predicting daily human behaviors. On the other hand, time-series forecasting is challenged by the uncertain and complex external environment, such as weather conditions that affect prediction accuracy. A promising method to improve the prediction accuracy is to adopt multiple external environment variables. Regarding this, the paper proposes using the multivariate dataset and the Kalman filter (KF) to predict the electrical power consumed by the smart home appliance. We conduct extensive experiments based on the real datasets of power consumption, which are classified into multivariate and univariate and used in the LSTM-KF model to predict the power consumption of the smart home appliance. The LSTM here stores the data information for static prediction, and the Kalman filter dynamically adjusts the prediction results to obtain a final prediction value. The LSTM-KF models applying the proposed multivariate and the univariate are compared in terms of the RMSE and the determination coefficient R2. The LSTM - KF using multivariate shows the best accuracy. Nonetheless, the univariate method using the Kalman filter outperforms the multivariate method without using the Kalman filter, implying the significance of using multiple variables together with the Kalman filter in improving the prediction accuracy.