Fusion of energy sensors with missing values

In Smart Energy Grids, the information flow used to make decisions is the result of fusion of different sources. Communication latency, possible sensor faults and inaccuracies, may negatively impact the data quality and hence the taken decisions. For these reasons, the construction of a robust representation of the input signals that replaces and/or corrects the inaccurate data is crucial for effective classification, anomaly detection and planning. Recent works on Data Fusion and data imputation suggest that the usage of other signals in the same context can empower the representation and can be a useful preprocessing task. In this work we describe an Autoencoder-based data fusion architecture with convolutional layers, skip connections and ad-hoc augmented training sets for data imputation applied to the power consumption measurements obtained by different sub-meters. Among the investigated architectures, the approach with the shared convolutional layers and an augmented dataset that consider missing data in the random positions and located in the central part (AE-A-ALL-CNN), is the most promising one. In presence of one half of the input signal, in the central part, completely erased, it improves the imputation capability, respect to two most employed approaches (denoising autoencoder and MICE) in the average of 12 %.