TransLearn: A clustering based knowledge transfer strategy for improved time series forecasting

摘要

The widespread usage of time series prediction has led to the formulation of several hybrid and statistical algorithms. These algorithms assume the availability of significant quantities of labeled training data and utilize historical data observations to estimate future patterns. However, the assumption may be invalid in developing countries like India, where data availability is minimal due to several reasons such as fewer sensor deployments, newly added sensor locations and many more. Hence, the data scarcity issue poses a significant challenge to the target model’s prediction accuracy and generalization capabilities. To address this issue, a transfer learning model integrating the clustering analysis with deep neural architectures (TransLearn) is proposed in the current research study. The clustering algorithm with Soft-Dynamic Time-warping measure creates time-series groups following similar characteristics, trends and seasonality patterns, thus, creating a larger dataset for deep neural models training. The performance estimation of the proposed TransLearn strategy is conducted on the novel energy consumption dataset of several northern states of India. It involves employing clustering to group states with similar load variation patterns. Subsequently, four transfer learning-based deep neural models are trained on the clustered datasets. Lastly, the fine-tuned results of these TransLearn models on each individual state are evaluated against the conventional deep neural model building strategy. The comparative evaluation in terms of widely used performance measures shows that the TransLearn approach outperforms the conventional neural models by generating the least prediction error.