Urban micro-climate prediction through long short-term memory network with long-term monitoring for on-site building energy estimation

Accurate meteorological data play a substantial role in the building energy estimation process and projected energy savings retrofitting. The present study presents predicted micro-climates parameters with long short-term memory (LSTM) network based on the long-term on-site measurement and its significance in the building energy analysis. The one-day-period-ahead prediction results demonstrated approving performance that the average RMSE of predicted on-site temperature is 0.75 °C, corresponding to 4.11% in MAPE while RMSEs of EPW data (the common embedded datasets representative of the typical meteorological year) and suburban meteorological station data are 5.23 °C and 5.18 °C, respectively; the similar applied to relative humidity and solar radiation. The predicted meteorological parameters were therefore passed into building energy estimation models. The comparisons of energy consumption for building heating and cooling against reference models with suburban station climates and EPW datasets are statistically investigated, with the underlying propagation of bias from meteorological inputs being analyzed. For the typical building where the micro-climate station located, the estimation biases are as follows (i) LSTM predicted datasets: Δ = -1.58% for cooling, Δ = -2.51% for heating; (ii) EPW climate datasets: Δ = -29.68% for cooling, Δ = +129.88% for heating; (iii) suburban station climate datasets: Δ = -5.1% for cooling, Δ = +235.95% for heating.