Using an artificial neural network model for natural gas compositions forecasting

The paper presents the results of natural gas composition forecasting obtained using the MLP model of artificial neural network. The training of MLP model was performed on the basis of 8760 real data describing the per-centage shares of the five main components of natural gas in a selected town on the territory of Poland. The model includes calendar factors (month, day of month, day of week, hour of day) and weather factors (ambient temperature), which indirectly affect the composition of natural gas in a given point of the gas network and were selected after a detailed statistical analysis of the variability of the natural gas composition in time. Based on the value of the correlation coefficient for the test set and the MAPE forecast errors calculated on the basis of the actual and the forecast data, the best quality MLP 18-65-5 network was experimentally selected. Natural gas composition forecasts were made using this model for input data characterizing the next calendar year and the average MAPE forecast error = 3.356% was calculated. Risk analysis, in turn, was used to estimate the proba-bility of obtaining a forecast with a MAPE error greater than the mean error.