Surrogate modeling for long-term and high-resolution prediction of building thermal load with a metric-optimized KNN algorithm

Abstract

• Surrogate modeling, to predict all-year hourly building loads in pre-design stage. • Load database contains 16384 models with crucial features and thermal load data. • Manhattan distance is the optimal metric when performing KNN in load prediction. • Propose thermal load prediction process at district level and apply it to a case. During the pre-design stage of buildings, reliable long-term prediction of thermal loads is significant for cooling/heating system configuration and efficient operation. This paper proposes a surrogate modeling method to predict all-year hourly cooling/heating loads in high resolution for retail, hotel, and office buildings. 16 384 surrogate models are simulated in EnergyPlus to generate the load database, which contains 7 crucial building features as inputs and hourly loads as outputs. K-nearest-neighbors (KNN) is chosen as the data-driven algorithm to approximate the surrogates for load prediction. With test samples from the database, performances of five different spatial metrics for KNN are evaluated and optimized. Results show that the Manhattan distance is the optimal metric with the highest efficient hour rates of 93.57% and 97.14% for cooling and heating loads in office buildings. The method is verified by predicting the thermal loads of a given district in Shanghai, China. The mean absolute percentage errors (MAPE) are 5.26% and 6.88% for cooling/heating loads, respectively, and 5.63% for the annual thermal loads. The proposed surrogate modeling method meets the precision requirement of engineering in the building pre-design stage and achieves the fast prediction of all-year hourly thermal loads at the district level. As a data-driven approximation, it does not require as much detailed building information as the commonly used physics-based methods. And by pre-simulation of sufficient prototypical models, the method overcomes the gaps of data missing in current data-driven methods.