A Modified Fuzzy Logic Relation-Based Approach for Electricity Consumption Forecasting in India

Abstract

Prediction of electricity demand is made using Load forecasting technique to meet the ever-growing demand. In this paper, future electricity demand forecasted for the whole state of Uttar Pradesh (India), using the dataset collected from the Central Electricity Authority. This dataset consists of electricity demand for the whole state of UP for every 15-min block. Different models were used to forecast future demand. XGradientBoost (XGBoost), a machine learning algorithm was used to forecast demand first. Further forecasting was performed using deep learning models such as Long Short-Term Memory (LSTM) using neural networks as they are considered to be more efficient and accurate than XGBoost. Fuzzy time series (FTS) models were considered to incorporate trend and seasonality present in our dataset. From various FTS models, the best mean absolute percentage error achieved (MAPE) was 2.34%. A new method KmFuzz is proposed in this paper that uses modified K-Means clustering for finding an optimal number of partitions on which fuzzy logic is applied. Fuzzy sets are obtained by applying fuzzification on the dataset and the total number of sets generated are equal to the number of optimal partitions. Then the weighted average method is used for defuzzification and forecasting the next hour demand using the demand data of previous hours with MAPE of 1.94%, thus improving the accuracy further.