Imputation of missing values for generating typical meteorological year (TMY) with data decomposition and recurrent neural networks

Abstract

Typical meteorological year (TMY) for a specific location is critical information when designing low-carbon and energy-saving buildings. However, in developing countries, long-Term observations of weather are now readily available and even mixed with missing values. In this study, a nonlinear autoregressive (NAR) recurrent neural network model in combination with improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method is demonstrated for treating the missing values in observed daily average air temperature at Bogor City in Indonesia. The prediction results for missing values indicate that the ICEEMDAN-NAR hybrid model performs very well with high accuracy when compared with the observed in the validation and model comparison phases. Particularly, the relatively lower frequency oscillation modes in observed data can be predicted well. Thus, this method can be used for relatively medium-and long-Term prediction of missing values with respect to the given data/input period. critical to consider cities and buildings from the eyes of older people in society.