TCN-BiLSTM-CE: An interdisciplinary approach for missing energy data imputation by contextual inference

Energy data frequently contain missing values, which complicate accurate analysis and informed decision-making. Since energy load data are generally in the form of time series, the problem of reconstructing missing data in energy load data can also be translated into the problem of reconstructing missing data in time series data. Inspired by image-based techniques, this study introduces a novel approach for imputing missing energy data using a Context Encoder (CE), in order to translate the original network from the task of reconstructing the missing parts in image data to the task of reconstructing missing data in time series data, using the complete data before and after the time period in which the missing data appears as the contextual reference to infer the missing data in the missing time period. Two structures commonly used in time series data processing, TCN and Bi-LSTM, were introduced to learn the latent representation in latent space, in order to extract the features of bidirectional contextual information, causal relationships and long-term dependency in time series data. After conducting separate experiments for electrical load data and heat load data, we demonstrate the effectiveness of our proposed method on energy data for different energy types. Compared to other baseline methods, the Normalized Root Mean Squared Error (NRMSE) between original and reconstructed electric load data and heat load data reduced from 7.58 %–81.20 % and 4.26%–65.74%, respectively, fully demonstrating the potential of using the proposed method for reconstructing the missing energy load data. • Cross-domain methods are used for image filling to adapt models suitable for filling missing values in energy demand data. • Introduction of Temporal Convolutional Network and Bi-directional LSTM enhances applicability to contextual information extraction of energy data. • The reconstruction loss and the adversarial loss are used during the network training for calculating the loss with and without missing values, respectively. • Missing data of different energy types, different types of missing values (continuous, discrete) in different periods were reconstructed and evaluated separately. • Model for missing energy data imputation adapted from image-filling model outperformed LSTM models, SOTA models and original image model.