Continual learning for time series data analytics

Abstract

Various traditional time-series forecasting models have been implemented in the past, including ARIMA, but they are limited in their ability to capture complex non-linear relationships and adjust to new data, leading to inaccurate predictions. The section also mentions some machine learning methodologies such as SVM that have shown promising results in HVAC load forecasting. This project aims to address the limitations of conventional time-series forecasting models by utilizing a Recurrent Neural Network (RNN) - Long-Short Term Memory (LSTM) model and implementing Continual Learning (CL) - incremental learning methodology to optimize the model's self-learning ability. The objective is to enhance energy consumption efficiency while maintaining the desired thermal comfort level for HVAC systems in Singapore. The project involves data pre-processing, applying the RNN-LSTM model, assessing the model's performance, optimizing it with CL, comparing forecasting results, and exploring interdependence among various HVAC system parameters. The RNN-LSTM model performed well on the training data and showed promise for accurate prediction. The incremental learning approach effectively integrated new data during each iteration, enhancing the model's performance. Finally, linear regression and Granger Causality analyses provided insights into the linear relationship between the "AHU" and "FCU" variables, with the coefficients and intercept values being useful for predictions and showing a significant causal relationship between the two variables at lag 1.