IN-PERSON: Livable Cities – New York

Abstract

The purpose of this study is to predict the solar power using ANN(Artificial Neural Network) and LSTM(Long Short Term Memory) technique. In the case of demand data(Period : Jan. 1st. 2019 to Dec. 31st. 2019), residential buildings were calculated through EnergyPlus, and other buildings were calculated based on measured data. In the case of supply data(Period : Jan. 2st. 2019 to Jul. 4st. 2019), it was calculated based on the measured data provided by Jincheon Eco-friendly Energy Town. PCC(Pearson Correlation Coefficient) was performed for 13 variables and 1 target value. The variable selection is that r square is more than 0.7 or absolute value r is more than 0.3. Among them, 6 variables satisfied the condition. The dataset was train data 70% and test data 30%. Epoch was set to 100. Because, as a result data analysis, if learning goes over 100 times the loss value does not change anymore. I analyzed 155 cases each with ANN and LSTM. The structure of the model with the best prediction performance are ANN, 3 Hidden Layers and Hidden Node of 14, 13 and 11 respectively. In addition, CV(RMSE) was 29.1% and NMBE was -7.14%, both of which satisfied the ASHRAE guidelines. The three cases were compared(Case 1: Demand energy, Case 2: Demand energy & Supply energy, Case 3: Demand energy & Supply energy & Energy storage system). The annual comparison results are as follows. Case2 decreased by 46.3% compared to Case1. And Case3 decreased by 56.3% compared to Case1.