Predicting Electricity Usage Based on Deep Neural Network*

Abstract

This paper describes a deep neural network (DNN) based method for forecasting short-term hospital electricity usage. In Experiment One, a 4-layer DNN stack auto-encoder (SAE) based model is constructed to verify the accuracy of the method. Kilowatt-hours (kwh), capacitance (pf), power factor (phi), voltage (v), electricity reactive power (var), and electricity active power (w) are the main input variables. After training the model, the prediction accuracy can reach 77.60%. In the improvement phase, the model is altered to use more common variables; specifically, kilowatt-hours (kwh), electric charge (charg), average active power (avg-w), and maximum active power (max-w) are used as input variables. In order to optimize the training of the model, Experiment Two improves on the basis of the original DNN model. As a result, the prediction accuracy can be increased to 85.17%. Finally, the four power data with the best measurement are used, namely current(I), voltage(V), reactive power(Var) and active power(W), and the predicted result is 98.14%. This method indicates that the planning and scheduling of the hospital’ s electricity usage will also be improved.