Previous studies with artificial neural network (ANN) application on air quality prediction show success even most models developed use only temporal data of air concentrations; hence mainly time series analyses are performed. It is known that the concentrations of air pollutants are highly related to the variations of the local and regional meteorology conditions which dictate the dispersion and transport routes of them. Many correlation studies and models simulating the fate and impacts of the released pollutants are these bases. Meanwhile, ambient air pollutants may also affect the concentrations of each other therefore making prediction or modelling of their behaviours a very complex problem. This study aims at designing economic and flexible ANN models for the 24-hour-ahead predictions on concentrations of respiratory suspended particulates (PMio) taking into account the effects from local meteorological conditions and related pollutants. The ANN applied in this study is a three-layer feed-forward network (TLFN) of the back-propagation type. To achieve computation efficiency, the size of the input data is limited to six parameters per input set. In addition, variation of the predicted hourly PMio concentration is assumed to depend only on meteorological and air quality conditions within the last 72 hours. With these as the main criteria, models for PMio concentration prediction are developed and tested with one year of hourly measured data in a small coastal city, Macao. The measured data include seven meteorological parameters (dew point, wind direction, wind speed, relative humidity, precipitation, sunshine hour fraction, and temperature) and concentrations of five criteria pollutants (PMio, SO2, NO/NO2/NOX, O3, and CO). Based on previous study, these data are divided into two seasons namely summer and winter due to the prevailing monsoon climate; hence the development of the PMio models for these two seasons are treated independently. Individual season segment of data is further divided into three sub-sets for model training, testing and validation purposes. To select the proper six input parameters for each seasonal model, correlation coefficients among the hourly concentrations of the listed pollutants and meteorological parameters are calculated. Six input parameters with the highest absolute values of correlation coefficients are selected to form the model input pattern with the minimum correlation coefficient value for cut-off being 0.73. The number of neurons used in the hidden layer for each model with its selected six-parameter input pattern is determined by systematic trials of minimum root mean square error. Five and six neurons are determined for the summer and winter models. The trained models are then used to predict PMio concentrations for seven days and compared with actual measurements. Absolute relative percentage error (ARPE) and mean absolute percentage error (MAPE) are used as comparison criteria. Results show that 75% of the predictions in both models achieve an accuracy of ARPE less than 30%, while 100% of the predictions in the winter model could achieve an accuracy of ARPE less than 40%. The overall MAPE for the two models are 49% and 77%, respectively. Overall, the developed ANN models can capture the general trends of the real measurement. Larger errors occur at those hours with high and sharp concentration peaks which may be considered as noises caused by instantaneous, individual incidents occurred in the vicinity of the monitoring station. These abnormal individual incidents may not be easily captured based on knowledge of the past.