Heating energy consumption prediction based on improved GA-BP neural network model

An improved genetic algorithm-back propagation (GA-BP) neural network model with fixed seeds, reasonable training sample size (TSS) and total number of nodes in hidden layer (TNNHL), as well as metabolism was established to predict heating energy consumption, thereby optimizing district heating (DH) operating strategies. Error variance, mean absolute percentage error (MAPE), and coefficient of determination (R2) were used as evaluation indicators. Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) and entropy weight methods were employed to comprehensively evaluate the prediction accuracy and acquire the optimal combination of TSS and TNNHL. A DH system in Beijing, China was used as the case study. Results show that the prediction results under fixed seeds are better and more stable compared with random initialization. Error variance, MAPE, and R2 vary between 0.0016 and 0.0202, 2.43% and 5.18%, and 0.590 and 0.967, respectively, when TSS and TNNHL change from 2 to 9 and 2 to 14, respectively. Error variance is positively correlated with MAPE and negatively correlated with R2. When TSS and TNNHL are both 2, the best prediction accuracy can be obtained, with error variance, MAPE, and R2 of 0.0016, 2.43%, and 0.967, respectively, verifying the accuracy of the proposed model.