A flexible fuzzy inductive reasoning approach for load modelling able to cope with missing data

Abstract

Load forecasting in buildings and homes has been in recent years a task of increasing importance. New services and functionalities can be offered in the home environment due to this predictions, for instance, the detection of potential demand response programs and peaks that may increase the energy bill in a dynamic tariff framework. Almost real-time predictions are key for these services but missing values can dramatically affect the performance of the energy forecasting or distort the prediction significantly. Fuzzy Inductive Reasoning has been proven to model load consumptions with high accuracy compared to other typical AI and statistical techniques. Nevertheless, it has several limitations when missing data is presented in the training data of the model and during prediction. In this paper, we present an improved version of Fuzzy Inductive Reasoning, called Flexible FIR Prediction that can cope with missing information in the input pattern as well as, in situations where patterns are not found in the behaviour matrix. The new technique has been tested with real data from one building of the Universitat Politècnica de Catalunya (UPC) and the results show that Flexible FIR Prediction is able to generate good predictions with low errors (less than 15%) although missing data is present in the training and online prediction phases.