Optimization of PV Systems Using Data Mining and Regression Learner MPPT Techniques

ABSTRACT


ISSN: 2502-4752 
Optimization of PV Systems Using Data Mining and Regression Learner… (Adedayo M. Farayola) 1081 summarization, regression, and association rule learning (dependency modelling) [12] .Cluster analysis is used to learn the similarities that exist in data sets.Anomaly detection detects the error in data sets.Classification simplifies the known structure in new data sets.Summarization offers an extra compact representation of the data sets while regression learning estimates the relationship with data sets and attempts to find a function that models the data sets with least error.Regression learning uses predictive analysis to obtain more accurate results from a decision support system.Dependency modelling examines the relationships with variables [13].
The contribution of this paper is to introduce the use of a particular class of data mining techniques known as regression learning algorithm for optimization and improvement of MPPT controllers in PV systems.The regression learning algorithms considered are linear support vector machine (LSVM) regression technique and rational quadratic gaussian process regression (RQGPR) technique.These regression algorithms use few real-time samples to generate the new data sets needed to train the ANN for MPPT task.Second contribution is a work done to evaluate the feasibility of the two proposed regression learning algorithms (LSVM and RQGPR) that were combined separately with ANN as ANN-LSVM and ANN-RQGPR technique respectively for the enhancement of MPPT techniques in PV systems under different weather conditions.
The synopsis of this paper is prepared as follows, section 2 will present a summary of the used MPPT techniques.In section 3, a report of the experiment setup and method is provided.Section 4 will present the results, and section 5 will include the conclusions.

MPPT TECHNIQUES
The techniques used in this study are briefly discussed:

Linear-Support Vector Machine (LSVM)
Support vector machine (SVM) is a popular machine learning technique used for classification and regression analysis.SVM regression technique was first introduced by Vladimir Vapnik in 1992 and has been widely used to solve pattern recognition problems, particle identification problems, and optimization problems [5].Examples of SVM regression techniques include the fine SVM, medium SVM, coarse-gaussian SVM, and linear support vector machine (LSVM) regression technique [13].The LSVM regression technique aims at finding a linear function f(x) with the minimal norm value (β'β) that makes the function f(x) to be as flat as possible using what is called primal formula and dual formula.The primal formula is briefly described using Equations (1-2), while Equations (3-6) illustrate the working principle for the dual formula used in linear SVM regression analysis, Where xn are the data sets comprising of N observations, yn is the response, εn and εn* denote the slack variables for each point, while αn and αn* are non-negative multipliers for each observation xn [14].

Rational Quadratic-Gaussian Process Regression
Rational quadratic gaussian process regression (RQGPR) is a type of gaussian process regression (GPR) technique that uses rational quadratic kernel (covariance) in solving optimization, regression, training and prediction problems [15].The regression kernel helps to determine the characteristics (e.g.fitness, smoothness, periodicity, etc.) of a function f(x).The RQGPR kernel is denoted using Equation (7).The RQGPR technique exhibits some similarities with the squared exponential GPR when the scale mixture parameter (σ) from Equation (7) approaches zero [16].The above-mentioned square exponential GPR kernel can be represented using Equation (8), Where K (x, x') is the covariance kernel, variable x and x' are the input variables, σ 2 is the maximum covariance, and l is the scaled-length used to determine how quickly a GPR varies with the input variable (x).

Artificial Neural Network (ANN)
ANN is a machine learning technique that works like the human brain and is used to solve both linear and non-linear tasks.ANN comprises of three layers; the input layer, hidden layer, and the output layer [19].The input layer receives the information (training data), processes the data through learning, and gives out predicted output data through the output layer.The hidden layer is an invisible layer with its output interconnected to the inputs of some other neurons [20].In photovoltaic systems, ANN input variables can be the irradiance (G), temperature (T), open circuit voltage (Voc) while the ANN predicted response can be duty cycle (D), predicted current, predicted voltage, or predicted PV power.The ANN neurons process, evaluate the input signal using linear method, then compare with its sum by means of a non-linear function known as activation function, and sends the consequence to other neurons.ANN has two common types, the feedforward neural network and the recurrent neural network [21].A neuron is modelled using equation ( 9), where the input variables are denoted by X 1 , X 2 , X 3 , …, X m and the respective weight of the individual inputs are denoted by W 1 , W 2 , W 3, …, W m respectively [5], [22].Mathematically, the neurons in the hidden layer can be estimated using equation (10), where N h is the hidden layer, N i is the input layer, and N o is the output layer.Some drawbacks with ANN technique include longer processing time for large networks, complexity of the ANN algorithm, and ANN procedure demands that the system is first trained using prior data sets and the collection of these training data sets might be cumbersome [23]- [25].

SIMULATION MODEL
To investigate the feasibility with the use of regression learning algorithms in tracking the maximum power point in a stand-alone photovoltaic system, an experiment was conducted using a complete photovoltaic system that includes a soltech 1STH-215-P PV panel, modified cuk DC-DC converter, MPPT controller, and a 20 Ω resistive load.The training data sets were collected from PSIM software.Table 1 illustrates the specifications of the used PV panel and the DC-DC converter in this study.The PV efficiency, load efficiency, and DC-DC converter loss were obtained using Equations (11)(12)(13) MCUK Losses input power -output power  (13) Where P pv(mppt) is the 1STH-215-P rated power at STC (standard test condition), P pv(max) is the PV extracted power, and P out is the output power at the 20 Ω resistive.For the third fold, Figure 2 displays a complete stand-alone PV system designed using ANN-LSVM technique.The ANN-LSVM output (reference current (Iref*)) was compared with the PV current (Ipv) as error signal (Iref* -Ipv).The error signal was passed through a power-integral (PI) controller for fine tuning and outputs the duty cycle signal (D).The duty cycle signal was transmitted through a pulse width modulator (PWM) as pulse signal that was used to activate the Mosfet gate of the DC-DC converter.Table 2 displays the average testing error of the trained ANN-LSVM MPPT technique using a total of 129 samples that comprise of 10 samples from the PSIM data sets and 119 samples from the generated ANN-LSVM data sets in the proportion 70% training, 15% testing, and 15% validation.
For the rational quadratic gaussian process regression (RQGPR) technique, similar procedures as shown in Figure 1, and a block diagram as shown in Figure 2 of a complete PV system but using RQGPR algorithm and ANN-RQGPR MPPT technique were used.Table 3 presents the ANN-RQGPR training, testing, and validation statistics.Where a value of R that is approaching 1.00000 and a MSE approaching 0.00000 validate a well trained model.For the conventional ANN technique, Figure 3 shows the complete PV system designed using a conventional ANN MPPT technique.Unlike the ANN-LSVM and ANN-RQGPR techniques that were trained using optimized data sets from the regression learning predictions, the conventional ANN algorithm

RESULTS AND ANALYSIS
Table 5 and Figures 4-9 present the tabulated results and the graphical results for the conducted experiment using conventional ANN technique, non-conventional ANN-LSVM and non-conventional ANN-RQGPR MPPT technique under three different weather conditions (NOCT, PTC, and STC).The NOCT is the normal operating cell temperature where the irradiance (G) is 800 W/m2 and the ambient temperature (T) is 47.40 °C.STC is the standard test condition where G is 1000 W/m2 and T is 25 °C while PTC is the PVUSA test condition where G is 1000 W/m2 and T is 20 °C For case 1 (NOCT), where G is 800 Wm-2 and T is 47.40 °C, obtained results show that the conventional ANN and the non-conventional ANN-RQGPR technique had tie (same) and best results at both the PV end and at the 20 Ω resistive-load end (73.09%PV efficiency and 69.79% output load efficiency) while ANN-LSVM displayed the lowest performance (72.67%PV efficiency, 69.43% load efficiency).However, the ANN-LSVM exhibited the lowest DC-DC converter power loss under NOCT condition.Figures 4-5 display the graphical results of the extracted PV power and the output load power from the PV system using conventional ANN, non-conventional ANN-LSVM, ANN-RQGPR technique, and under NOCT weather condition.
For case 2 (PTC), where G is 1000 Wm-2 and T is 20 °C, ANN-RQGPR had the best result (101.95%PV efficiency and 97.71% resistive load efficiency), while ANN-LSVM underperformed (100.40%PV efficiency and 97.71% load efficiency).Similarly, the DC-DC converter loss in power with ANN-LSVM regression technique was the lowest (8.56 W).Figures 6-7 display the graphical results of the extracted PV power and the output load power from the PV system using ANN, ANN-LSVM, ANN-RQGPR technique, and under PTC weather condition.
For case 3 (STC), where G is 1000 Wm-2 and T is 25 °C, both ANN and ANN-RQGPR achieved close results at the PV end as equal powers and efficiencies were extracted from the PV system (212.90W PV power and 99.88% PV efficiency) whereas the PV panel did underperform using ANN-LSVM (209.90W PV input PV power and 98.48% PV efficiency).However, at the 20 Ω resistive end, ANN-RQGPR overperformed by extracting the maximum output power and efficiency (204.16W resistive-load power and 95.78% load efficiency).Also, the ANN-LSVM MPPT technique underperformed as 201.45W power was produced at the resistive load end and with a load efficiency of 94.45%.Figures 8-9 display the graphical results of the extracted PV power and the output load power from the PV system using ANN, ANN-LSVM, ANN-RQGPR technique, and under PTC weather condition.
In addition, at STC, the ANN-LSVM had the lowest DC-DC converter power loss as 8.45 W was dissipated at the DC-DC converter while power dissipated with ANN-RQGPR MPPT technique in all the three environmental condition cases (NOCT, PTC, and STC) were the highest.Also, from the analysed regression (R) and the mean-square-error (MSE) results shown in Tables 2-4

CONCLUSIONS
This paper presents an innovative use of a particular type of data mining technique known as the rational quadratic gaussian process regression (RQGPR) learning algorithm to track the maximum power point in PV systems.Findings suggest that optimization of PV systems using RQGPR technique can be used to extract maximum power from a photovoltaic panel under different weather conditions.The RQGPR can successfully generate the large, acceptable and accurate training data sets needed to train the supervised machine learning techniques like ANN and ANFIS for MPPT tasks.Also, the mean-square-error (MSE) and regression (R) error statistics with RQGPR technique were better than that of conventional ANN and linear support vector machine (LSVM) regression techniques.Results confirmed that RQGPR technique exhibited an improved tracking power and efficiency compared to the linear support vector machine (LSVM) regression technique.

Figure 1
Figure1presents the flowchart algorithm of the linear support vector machine (LSVM) regression technique.The LSVM modelling was done in three folds.The first fold dealt with the optimization and generation of a fitness function (yfit) using LSVM regression kernel and few collected samples of the PSIM data sets(19 instances)  in training the model.The data sets comprise of two input variables (different levels of irradiance(G) and temperature (T)) as predictors (X) and one output variable (reference current (Iref* )) as the response.The LSVM regression kernel was then used to predict the responses (Iref) for an additional 110 instances comprising of variables G and T. The second fold dealt with the training, testing, and validation of the LSVM newly predicted data sets using ANN technique.The LSVM generated data sets were split in the proportion 75% for training, 15% testing, and 15% validation.For the third fold, Figure2displays a complete stand-alone PV system designed using ANN-LSVM technique.The ANN-LSVM output (reference current (Iref*)) was compared with the PV current (Ipv) as error signal (Iref* -Ipv).The error signal was passed through a power-integral (PI) controller for fine tuning and outputs the duty cycle signal (D).The duty cycle signal was transmitted through a pulse width modulator (PWM) as pulse signal that was used to activate the Mosfet gate of the DC-DC converter.Table2displays the average testing error of the trained ANN-LSVM MPPT technique using a total of 129 samples that comprise of 10 samples from the PSIM data sets and 119 samples from the generated ANN-LSVM data sets in the proportion 70% training, 15% testing, and 15% validation.For the rational quadratic gaussian process regression (RQGPR) technique, similar procedures as shown in Figure1, and a block diagram as shown in Figure2of a complete PV system but using RQGPR algorithm and ANN-RQGPR MPPT technique were used.Table3presents the ANN-RQGPR training, testing, and validation statistics.Where a value of R that is approaching 1.00000 and a MSE approaching 0.00000 validate a well trained model.

Table 1 .
The Used PV Panel and MCUK DC-DC Converter Specifications

Table 2 .
The Regression (R) and Mean Square Error (MSE) Statistics for ANN-LSVM Technique

Table 5 .
Results of the conducted experiment under varied weather conditions