Solar radiation forecasting with multiple parameters neural networks
Introduction
The best solutions for alternative energy sources, such as some of the solar energy and photovoltaic (PV) systems are rapidly gaining acceptance [1]. Solar water heating, photovoltaic, energy balance studies of the atmosphere, agricultural science and weather forecasting are implemented in various fields of solar radiation [2]. The database of climate with different empirical models are developed, which are closely related to solar radiation as networks of meteorological stations provide reliable long-term climate data [3]. It could be described as a parametric (Prescott angstroms) and a nonparametric (ANN) based on mathematical forms. It (parametric and nonparametric) can estimate the solar radiation with high precision when well calibrated [4], [5].
Since modeling of solar radiation has been used for areasin various latitudes and with different climates, most countries of the world used ANN in the recent years [6]. ANN was also used for areas with different internal topologies different input variables (geographical and climatic) and multiple time scales (hourly, daily and monthly) [7]. As known set does not require input assumptions, the ANN approach has several advantages over conventional semi-phenomenological or empirical models [8], [9]. Review is to present an overview of the AI-techniques (expert systems, artificial neural networks, genetic algorithms and fuzzy logic and various hybrid systems) as a design tool for the optimal sizing of PV systems [10].
Artificial Neural Networks (ANNs) are used in the modeling of solar radiation worldwide based on locations and different climates. Most countries such as Oman, Spain, China, Greece, India and the UK are also working on the related areas [6]. In most applications during regression of neural network, the most difficult question that constantly rises is how to fix the number of delays and neurons in each layer. Although many years have passed there exists no exact technique that can respond to this query. Delay has a deep impact on the regression results based on its selection on network architecture. Conventionally, the architecture of ANN has been designed on number of trial and error basis, but some time it follow experimental [11], [12]. The time delay is mostly used in the training of neural network due to limitation in speed of data processing and variable nature of radiation. Most of the time delays cause fluctuation and uncertainty of the neural network. Hence, delayed network stability analysis has more importance on applied and hypothetical prospective [13]. Therefore, an artificial network has been raising attention on delay based results. The delay has more impact on the neural network and network stability which mostly depends on the number of delays [14]. Therefore, it is important for the network to reach global stability at an equilibrium condition [13]. Further, it is easier to optimize the analytic problem with good convergence rate and accuracy with delayed network, so that network capable to solving real time problem [15]. Therefore, several methodologies have been used for network in a constructive or else destructive way but common approach of solving problem uses cross-validation and early stopping techniques to decide the number of optimal delay [16], [17], [18].
The contribution of a neuron is significantly important in oversetting and under fitting during planning of the artificial network. Therefore the optimum neurons are fixed after analyzing multiple groups of neuron on minimum error criteria for proper convergence of the network. The architecture of neural network has become a research challenge for researchers; they tried and suggested several methods for fixing the neurons. Shuxiang et al., Osamu,.Tamura et al., Zhange et al., and Jin Yan Li et al., have examined novel method, numerical approximation, Akaike information criteria, set cover procedure and approximation number of neurons for network architecture, respectively [12], [18], [19], [20], [21], [22], [23].
One of the main unresolved problems of ANN is the lack of consensus on how to best implement them. The transfer function selection provides nonlinear mapping potential while modeling of natural phenomena in this section targets such question. The selection of transfer functions has no theoretical background so we compare different ANNs to seek better multistep ahead solar radiation forecasting performance [24]. The most desirable feature of ANN is a performance using activation functions [25]. Jordan, Liu, Sopena et al. obtained the logistic function, Gaussian basis and sine activation function ability to enhance a number of test parameters using neural networks [12], [18], [26], [27], [28], [29].
The article has used different parameters as inputs of the ANN model (latitude, longitude, altitude, location, rainfall, month, mean of maximum temperature, skin temperature, mean of minimum temperature, soil temperature, mean relative humidity, mean vapor pressure, total precipitation, cloudiness, mean of wind speed, mean duration of sunshine, date and month of the year). The results obtained by the ANN model were compared with the actual data, with the use of these input variables and error values were found within the acceptable limits [2], [3], [5], [30]. The literature reviews on the estimation of global solar radiation over India and explains the world for various meteorological parameters [23]. Furthermore, other parameters such as Diffuse Horizontal Irradiance (DH), direct normal radiation (DN), bulb point temperature (BP), due point temperature (DP), pressure (PR), wind direction (WD), wind speed (WS), and time in hours (HH ) number of Day (DD) and the number of months (MM) are equally important for radiation forecasts. But it is also necessary to focus on individual parameters with their own delay, neuron and TF calculation including other models. Alam, S. and S. Kaushik et al. introduce an artificial neural network (ANN) model, which is used for estimating beam solar radiation based on reference clearness index. The ANN technique is determined as a multivariable parameter such as latitude, relative humidity, rainfall, longitude, altitude, mean duration of sunshine per hour and months of the year [31].
The main objectives of this study were to determine the most appropriate independent variables of hourly global radiation in India. To develop an accurate model that varies from delay one to thirty, neuron starts from 10 at interval of 10 up to 300 and similarly eight different transfer functions are used for the most appropriate independent variables. This paper considers eight different neural network models for comparison and those are feed-forward network (FFN), Feed-forward back propagation (FFB), radial basis neural network (RBN), probabilistic neural network (PNN), Elman backpropagation neural network (ELM), cascaded feedforward backpropagation (CFB), custom network-1 (C1) and custom network-2 (C2). In particular, six of them are known and most used models, last two models are custom models.
Section snippets
Data processing and neural networks
We use the information provided by “Sunny” satellite on the website of the Indian Ministry of Renewable Energy (MNRE). The hourly data are collected; 5000 h of the year 2008 applied at the point 31.7069°N, 76.9317°E, and coordinates in India. Multiple parameters like GHI, DNI, DHI, BPT, DPT, PR, WD, WS, HH, DD and MM are also used at same point of time. ANN training was performed using the Levenberg-Marquardt algorithm (LMA), a difference of Newton׳s method to reduce tasks and the sums of the
Model evaluation and model behavior
To evaluate the ability of the model in two parts, a set of input variables belongs to learning processes and second data for the validation set. This validation sets are used as test sets. Test processes involve comparison of different model results. Finally, the model is selected based on the lowest root mean square error (RMSE) forecast error.where represents the measured value at the forecast horizon, is the predicted value and n is the total number of samples.
Conclusion
In summary, most of those specified models that had completed the test are not very convincing. However, papers published in reputable journals and their proposed models are reportedly very successful. Using large ANN forecasting model might work properly after recognize over-parameterization and over-fitting affect. Further study suggests the performance of large neural model is required before providing the same conclusion. The precise principles should be assumed to cover the experimental
References (63)
- et al.
Estimation of global solar radiation using ANN over Turkey
Exp Syst Appl
(2012) - et al.
Estimation of monthly-mean daily global solar radiation based on MODIS and TRMM products
Appl Energy
(2011) - et al.
Generation of synthetic daily global solar radiation data based on ERA-Interim reanalysis and artificial neural networks
Energy
(2011) - et al.
Artificial intelligence techniques for sizing photovoltaic systems: a review
Renew Sustain Energy Rev
(2009) Statistical estimation of the number of hidden units for feedforward neural networks
Neural Netw
(1998)- et al.
Bounds on the number of hidden neurons in three-layer binary neural networks
Neural Netw
(2003) - et al.
An integrated artificial neural networks approach for predicting global radiation
Energy Convers Manag
(2009) - et al.
Computation of beam solar radiation at normal incidence using artificial neural network
Renew Energy
(2006) - et al.
Forecasting of preprocessed daily solar radiation time series using neural networks
Sol Energy
(2010) - et al.
An adaptive wavelet-network model for forecasting daily total solar-radiation
Appl Energy
(2006)
Novelty detection: a review – Part 2: neural network based approaches
Signal Process
Predictive control of an integrated PV-diesel water and power supply system using an artificial neural network
Renew Energy
Estimation of monthly average daily global solar irradiation using artificial neural networks
Sol Energy
Use of neural nets for dynamic modeling and control of chemical process systems
Comput Chem Eng
A sequential learning approach for single hidden layer neural networks
Neural Netw
Multiple linear regression and artificial neural networks based on principal components to predict ozone concentrations
Environ Model Softw
Solar radiation estimation using artificial neural networks
Appl Energy
Artificial neural networks in renewable energy systems applications: a review
Renew Sustain Energy Rev
Solar radiation: cloudiness forecasting using a soft computing approach
Artif Intell Res
Neural network ensemble-based solar power generation short-term forecasting
JACIII
How many hidden layers and nodes?
Int J Remote Sens
Global asymptotic stability of recurrent neural networks with multiple time-varying delays
IEEE Trans Neural Netw
Convergence of discrete-time neural networks with delays
Int J Qual Theory Differ Equ Appl
Global exponential stability of bidirectional associative memory neural networks with time delays
IEEE Trans Neural Netw
Feedforward neural network construction using cross validation
Neural Comput
Cited by (94)
Solar radiation forecasting using gradient boosting based ensemble learning model for various climatic zones
2024, Sustainable Energy, Grids and NetworksBack propagation neural network based proportional-integral hybrid control strategy for a solar methane reforming reactor
2024, International Journal of Hydrogen EnergyDevelopment of a hybrid computational intelligent model for daily global solar radiation prediction
2023, Expert Systems with ApplicationsApplications of machine learning techniques in performance evaluation of solar desalination systems – A concise review
2022, Engineering Analysis with Boundary ElementsImproving the accuracy of predicting the performance of solar collectors through clustering analysis with artificial neural network models
2022, Energy ReportsCitation Excerpt :The application of such a combination of artificial intelligence (AI) techniques for solar thermal technologies represents a novel approach and an effective new way for predicting the performance of solar energy systems. ANN has been applied in solar energy, e.g., for solar radiation prediction (Vakili et al., 2017; Mghouchi et al., 2019; Bou-Rabee et al., 2017; Shaddel et al., 2016; Kashyap et al., 2015), photovoltaic applications (Hussain et al., 2020; Yadav et al., 2018; Almonacid et al., 2017) and solar drying (Prakash et al., 2016). ANN is also the subject of this paper.