Research On Building’s Electricity Consumption Prediction Using Data-driven Method

Building energy system is a kind of complicated system with multivariable and distributed parameters. Predicting building energy consumption level accurately is the important basis and premise to analyze the energy conservation potential of building and guide the energy utilization in the future. Meanwhile, it is beneficial to improve equipment’s efficiency of building energy system and reduce the waste of energy. In recent years, building energy consumption prediction using advanced optimization algorithm has received considerable attention. For existing building’s electricity energy consumption forecasting, various advanced swarm intelligent optimization algorithms are applied in this paper. The specific work is as follows:(1) ANNs, a representative data-driven modeling method, have been widely applied in the field of building energy consumption forecasting in the past twenty years. This paper verifies the building energy consumption prediction model based on BP neural network. In the process of data preprocessing, a variable selection method based the PCA is used to reduce the dimension of input variables provided by ASHRAE. Simulation experiment results demonstrate that BPNN model can accurately predict building energy consumption, and the modeling time is short.(2) As the typical swarm intelligence algorithms, PSO and GA have been generally used for optimization issues. In order to overcome the shortcomings of BPNN, such as getting into the local extreme easily, converging slowly etc., PSO and GA are introduced to optimize weights and thresholds of BPNN. Experiment results indicate that PSO-ANN is superior to GA-ANN in modeling time, forecasting precision and computation complexity.(3) In view of the single PSO’s characteristics of premature convergence and slow search speed, an improved PSO algorithm combines the biological breeding and genetic variation mechanism is proposed. Breeding mechanism can effectively guarantee the preferable particle search in the solution space and accelerate convergence speed significantly. Genetic variation mechanism is used to disrupt the optimal trajectory of particles, which contributes to overcome the phenomenon of fall into local optimal solution easily. The benchmark functions test results show that iPSO has better optimization performance and faster global convergence speed. Then, a novel building energy consumption prediction model is established based on improved PSO and ANN(iPSO-ANN). Prediction results indicate the average modeling time of iPSO-ANN model is in less than 10 seconds. Prediction accuracy is higher than that of single BPNN by 22.7%.(4) The hourly building electricity energy consumption is predicted by proposed iPSO-ANN model. Building energy consumption data are acquired from the campus energy consumption monitoring platform and the relevant meteorological data from local meteorological agency official information. The experimental results indicate that iPSO-ANN model has the strong reliability of short-term building electricity energy consumption, which can replace PSO-ANN, GA-ANN and ANN models to realize the online building’s electricity energy consumption forecasting.