Evaluation and Comparison of FFB, CFB, and MLP Artificial Neural Networks for the Identification of Potential Sites for the Construction of Photovoltaic Solar Power Plants in East Azarbaijan Province

As technology evolves in many countries around the world, the need for energy is increasing, which is especially important in developing countries such as Iran, because sustainable energy is needed to develop the process of sustainable development. Due to the geographical location of Iran and having a large number of sunny days, using solar power at the scale of the power plant helps provide sustainable energy. According to the radiation map provided by the Iranian New Energy Organization in East Azerbaijan province, there is enough potential to build a solar power plant. Due to the ability of neural networks to solve complex problems, the present study has used a combination of spatial decision-making system, GIS environment, and artificial neural networks to identify potential areas for solar power generation. The data used in the study include solar radiation, precipitation, sunshine hours, temperature, altitude, slope, LULC, distance from roads, and distance from cities. Based on these criteria, training data were obtained and trained using the Levenberg-Marquardt training algorithm of FFB, CFB, and MLP networks.According to the results, the CFB network, with form 9,6,1 and RMSE values 0.084 and 0.061 for training and test data, was selected as the most suitable network and with the results obtained from this network, the location was determined. The results were classified into five classes, with about 15% identified as very favorable for the construction of photovoltaic solar power plants in East Azerbaijan Province.

In this research, we are looking for zoning of photovoltaic solar power plants using an artificial neural network in East Azarbaijan province. Since artificial neural networks need training data to perform the calculations, the criteria are weighted first through ANP, and then by using the weights obtained for the criteria, the training layer for network training is created. Using the training layer, all three FFB, CFB, and MLP neural networks have been trained to obtain the appropriate network and optimal structure.Environmental criteria are selected based on the parameters of the construction of photovoltaic solar power plants. Given that the locating process is a multi-criteria decision problem between different parameters and criteria, therefore, the software must be selected that supports both the vector model and the raster model. It also can implement multi-criteria decision-making rules. Based on this, ArcGIS 10.6 software was used for data preparation, layer preparation, and integration. Super decision and Matlab software have also been applied to the process of analyzing network decision making and artificial neural networks.

The structure of neural networks is such that by changing the number of hidden layers and its neurons, the change of the stimulus function and training algorithm of the network structure is changed and affects the output of the model. Therefore, determining the optimal structure of the network is based on trial and error, and using the evaluation criteria and comparing the results, the optimal model is modeled with the least error. However, we should be careful that if the error rate is very close to zero in the evaluation of the training results, there is a possibility of over-fitting, which means that the network created will only be suitable for the training set and adding new data will not yield a satisfactory answer.Matlab software was used to simulate the structures of different artificial neural networks and determine the optimal structure. For the present study, three FFB, CFB, and MLP neural networks with different structures have been created so that all three networksemploy the Lonberg-Marquardt training algorithm with back-propagation error (trainlm). The number of neurons ranged from 1 to 15 and the number of repeats between 10 and 700. For the FFB and CFB networks, the tansig and purelin transfer functions and for the MLP network, the hardlim and hardlims transfer functions are investigated.According to the simulations, the optimal CFB network structure is 9,6,1 with 9 input neurons and 6 middle neurons, with MSE and RMSE values for the training data 0.006, 0.084 and for the test data 0.004, 0.061, the optimal FFB network structure as 9,5,1 with 9 input neurons and 5 middle neurons, with MSE and RMSE values for training data 0.11, 0.107 and for test data 0.012, 0.111 and the optimal structure of the MLP network as 9,9,1 with 9 input neurons and 9 middle neurons, with MSE and RMSE values for the training data 0.007, 0.085 and for the test data 0.006, 0.079 have been selected. Based on these results, the CFB neural network with the structure of 9,6,1 has the best performance among the networks. For this reason, the photovoltaic solar power plants in East Azarbaijan province have been located with this network.The final map was classified into five descriptive classes using the results obtained. According to the classification, about 7.7% were in the very undesirable class, 28.4% in the undesirable class, 27.6% in the middle class, 20.6% in the desirable class, and 15.5% were in the very desirable class.

With the advancement of industry and the development of new technologies, population growth in many countries of the world has increased the consumption of electricity. Also, all developed and developing countries have realized the fact that to maintain their international status, they need to provide sustainable energy, especially electricity, from non-fossil energy sources. As societies become more aware, the limitations and harms of using fossil fuels have become more apparent, forcing countries to source some of their electricity needs from other energy sources, such as renewable energy sources. Iran, like all developing countries, is no exception. Due to the geographical location of Iran and having 300 sunny days, the use of solar energy in both large and small sectors contributes to sustainable energy supply.In this study, the authors have tried to combine the existing methods for location namely the use of spatial decision-making systems and GIS, to use new methods such as artificial neural networks to identify potential areas for the construction of photovoltaic solar power plants in East Azarbaijan province. To accomplish this, based on the criteria for the construction and location of photovoltaic solar power plants, environmental factors include solar radiation, precipitation, sundial and temperature as climate criteria, elevation and slope as physical and land use criteria, distance from roads, and distance from cities are considered as economic criteria. Based on these criteria, training data was obtained through ANP, and along with this data, and LM training algorithm was performed to train FFB, CFB, and MLP networks.Based on the MSE and RMSE evaluation criteria, the CFB network with the structure of 9,6,1 was selected as the most appropriate network and the results were obtained from this network. After preparing the final map, it was determined that solar photovoltaic power plants could be built in the province.