Solar Irradiance Prediction in Central Africa with MLP Neural Network

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Feb 16 2024

In an article recently published in the journal Scientific Reports, researchers proposed a multilayer perceptron (MLP) neural network to optimize solar irradiance prediction in Central Africa using meteorological information as inputs.

Importance of reliable solar energy

The ongoing energy shortage in Central Africa is hindering sustainable development and long-term economic growth. Thus, renewable energy sources, specifically solar energy, have gained significant attention in this region as an effective solution to tackle this energy shortage. For instance, studies have demonstrated the substantial potential of solar energy and its critical role in facilitating sustainable energy development in Cameroon. Accurate solar irradiance predictions are crucial for developing sustainable energy plans.

Specifically, accurate solar irradiance projections improve the dependability and efficiency of solar energy production, which can increase the resilience and stability of the energy infrastructure of Central Africa. However, the estimation of solar irradiance is a difficult task due to the influence of several astronomical and geographical factors. The continuous interaction of atmospheric and meteorological conditions also increases the challenges of estimating solar irradiance.

Despite these difficulties, multiple estimating methods have emerged using meteorological data at various time intervals, such as monthly, daily, and hourly, to enhance solar irradiation forecasting accuracy.

The proposed approach

In this study, researchers assessed the feasibility of using an MLP neural network for forecasting solar irradiance on an inclined surface using meteorological information/climatic variables as inputs. MLP is an extensively researched non-recurrent artificial neural network (ANN) paradigm that provides a high flexibility in forecasting as it can accommodate different numbers of output and input variables.

The objective of the study was to use neural network methods to forecast hourly solar irradiance, which is critical for electricity production from photovoltaic sources. Researchers developed an ANN model by evaluating and training multiple MLP designs for solar irradiance prediction in the city of Douala based on climate parameters.

Multilayer back-propagation neural networks were programmed and designed using different architectures. The Douala Institute of Technology (IUT) meteorological station database provided 23 months of meteorological data, starting from January 2019 to November 2020, for testing, validating, and training the network.

Data was collected at 30-minute intervals during this duration. Subsequently, the collected data was saved in a database that underwent pre-processing in Excel. 15%/2360 samples of the available data were used for validation, 5% was for testing, and the rest 80%/11,024 samples were used for training the developed neural network model/MLP.

Four meteorological parameters, including atmospheric pressure, relative humidity, temperature, and wind speed, and two temporal parameters, including hour and day, were used as inputs to the network, while the solar radiation intensity was the network output.

Two performance metrics, including the coefficient of determination and mean square error, were employed to evaluate the effectiveness of the developed ANN algorithms. An ANN containing three layers, including one output layer, one hidden layer, and one input layer, was selected after several iterations.

The sigmoidal tangent (tansig) transfer function was used for the hidden layer, while the linear transfer function (purelin) was utilized for the output layer. Additionally, the Levenberg-Marquardt backpropagation algorithm was employed to train the ANN model.

Significance of the study

A logistic Sigmoid function with 50 hidden layer neurons was determined to be the optimal neural network architecture to compute the solar radiation intensities after obtaining the results from the extensive experiments performed using different input configurations.

This finding provided novel insights into the most efficient neural network arrangement for solar radiation forecasting under comparable environmental conditions. Regression analysis was performed to compare the predicted and target values to evaluate the MLP performance. In the analysis, the target was considered the independent variable, while the output was the dependent variable. The correlation coefficient was utilized as a metric that indicated the degree of proportionality between the targets and the output. A correlation coefficient close to unity was desirable to realize the optimal performance of the MLP.

The wind speed, relative humidity, atmospheric pressure, and temperature displayed correlation coefficients of 0.81224, 0.84381, 0.48392, and 0.97831, respectively. Remarkably, the MLP demonstrated a strong correlation coefficient of 98.83% between the observed and predicted solar irradiance, which provided empirical support regarding the efficacy of the proposed methodology.

Conclusion

To summarize, the study's findings showcase the effectiveness of the proposed MLP neural network approach in accurately predicting irradiance levels that almost match the true values. The proposed model can be feasibly used to calculate solar irradiance intensities at the Douala site and in other locations with comparable climatic circumstances.

Such high correlations are advantageous for implementing and strategizing solar energy programs in the Central African region. However, more research is required to investigate the model’s ability to adapt to various geographical locations and environmental conditions.