From time series to image analysis: A transfer learning approach for night setback identification of district heating substations

Abstract

District heating plays a dominant role in the heating markets of Nordic countries. Therefore, energy efficiency of district heating systems is of great interest to energy stakeholders. However, it is not uncommon that district heating systems fail to achieve the expected performance due to inappropriate operation in the buildings that they supply. Night setback is one control strategy that has been proved to be an unsuitable setting for well insulated modern buildings in terms of both economic factor and energy efficiency. In this study, a transfer learning approach is proposed for night setback classification of district heating substations using pre-trained convolutional neural networks as the main building blocks. In order to take advantage of the state of art performance by pre-trained models for computer vision tasks, a new way of problem formulation is proposed that converts the original time series data into heatmap images, which shifts the original research problem from conventional time series analysis into the image classification domain. The proposed method also makes it flexible and easy for the domain experts to switch back to manual examination even if the model fails. To evaluate the effectiveness of the proposed approach, hourly data of 133 substations in Oslo are used in the case study. Precision, recall, f1 score and accuracy are used as the performance measures. Results show that the overall performance of all models is reasonably good, with an f1 score greater than 0.9 and accuracy greater than 0.96 for the case where the models were trained with an imbalanced and a relatively small percentage of data using the proposed approach. • Approach works well for detecting night setback in district heating substations. • Framing the research problem using the approach offers great flexibilities. • Transfer learning can be adapted to solve time series analysis problems. • Transfer learning works well for imbalanced and small amount of training data. • Approach works well for a large amount of district heating substations.