Short‐term prediction of photovoltaic power generation based on neural network prediction model

Real‐time monitoring and accurate prediction of photovoltaic (PV) power generation operation parameters are essential to ensure stable operation. In this paper, a set of online PV power generation parameter measurement and monitoring devices characterized by simple structure, high sampling accuracy, small data fluctuations, and ease of measurement, are designed. Sensors based on the zero‐flux principle are employed in the real‐time collection of the output electrical signals in the process of PV power generation, realizing the accurate collection of electric parameter signals. Next, the basic structure and working principle of PV cells are analyzed, a mathematical model of PV cells for engineering purposes is established, a wavelet neural network is selected to predict the short‐term PV power generation, and particle swarm optimization and adding momentum are used to optimize the weight of wavelet neural network (WNN) as well as the parameters of the wavelet basis function. Finally, the historical power generation data and meteorological data of the power station are taken as the training samples to train and simulate the prediction sub‐models of different weather types to verify the effectiveness and accuracy of the PV power generation short‐term prediction model for optimizing the WNN based on the particle swarm optimization algorithm. The research results of this paper can realize real‐time monitoring of the output parameters and accurate prediction and evaluation of power generation during the operation of the PV‐power‐generation system.


| INTRODUCTION
Solar power generation has the advantages of wide resource distribution, economical and environmentally friendly properties, and sustainable development.
Therefore, solar energy is recognized as the ideal form of alternative energy and will have strong market competitiveness in the future of the global power sector. The principle of photovoltaic (PV) power generation is based on the conversion of sunlight irradiating the surfaces of the photocells into electrical energy via the PV effect. This electrical energy is converted via controllers and inverters into universal electricity and is transmitted into the power grid. Compared with power generation from conventional energy, the output power of the PV power generation system is restricted by environmental factors and its peculiarities. The environmental factors are uncertain and difficult to predict, resulting in the possibility of chance fluctuations in power generation and intermittent power outages of the system. For a power system, the main challenges of PV power generation and grid connections are variability and uncertainty. The output power of PV power generation may be variable and difficult to predict at different time scales. [1][2][3] Therefore, during the operation of a PV power generation system, monitoring real-time output parameters and conducting accurate prediction and evaluation of power generation is essential to ensure the operational quality of a PV power station. Establishing a prediction model with an appropriate algorithm and training the prediction model using the power generation and meteorological data of the PV power generation system is crucial to predict the output power generation in a future period of time. The statistical method in PV power generation involves conducting mathematical, statistical calculations on the law between the model input and the output to make further predictions. Neural network, 4,5 support vector machine (SVM), 6 and autoregressive-moving-average model (ARMA) 7 are the most widely used learning methods for data statistics. Bouzerdoum et al. 8 proposed a new short-term power prediction hybrid model for PV grid-connected plants. It combines seasonal autoregressive integrated moving average (SARIMA) and SVM, which is called SARIMA-SVM. They used a small power station database to experimentally verify the accuracy of the proposed model. Xu et al. 9 used a weighted support vector machine (WSVM) to predict short-term PV-power generation by choosing 5 days highly similar to that of the predicted day as the training samples and designing sample weights based on the similarity and time interval. The results showed that the WSVM-based output power prediction is more effective than artificial neural network-based output power prediction. Olatomiwa et al. 10 proposed a new method of mixing SVM with the firefly algorithm to predict the monthly average global horizontal irradiance with the sunshine duration, maximum temperature, and minimum temperature as predictors. To identify the weather types missing in the historical data, a weather pattern recognition model based on solar irradiation feature extraction and a support vector machine was proposed for the shortterm prediction of PV-output power. The results show that the developed support vector machines with firefly algorithm (SVM-FFA) model provides more precise predictions compared to artificial neural networks (ANN) and genetic programming models.
Neural network methods have been widely used in the PV power generation field because of their superior nonlinear fitting capability. 11 Minh-Thang et al. 12 utilized two sets of PV devices and used three different prediction models, such as the scalable sustainability model, an artificial neural network, and the multielement polynomial model. The results showed a long training time in temperate climates and a good performance of the trained model. Maria et al. 13 proposed wavelet decomposition and principal component analysis methods to decompose the meteorological data used as the prediction input and combined the least square SVM and time sequence forecast methods to predict meteorological data and realize the prediction of the PV-power generation on the same day. The results show that the data-driven forecast method combined with the data preprocessing techniques can improve the forecast accuracy, and decrease mean error bias by up to 1%. Lu et al. 14 established a radial basis function neural network with decoupling methods for the prediction of PV-power generation by comparing the PV powers obtained from the use of autoregressive integrated moving average (ARIMA), back propagation neural network (BPNN), and radial basis function neural network (RBFNN). The results showed that the proposed model could provide accurate and effective predictions of PV-power generation. Vishal et al. 15 constructed a hybrid neural network model of maximal overlap discrete wavelet transform, SARIMA, and random vector functional link to measure numerical predictions with precision. The results showed a high accuracy of the proposed prediction model. Mellit et al. 16,17 proposed an adaptive wavelet neural network method to predict the hourly power generation of an urban 20-kW grid-connected PV-power station. The prediction results showed that the model could achieve certain prediction accuracy under different weather conditions.
In related work, literature 18 presented a proposed deep learning method for PV prediction that combines autoencoder and long short-term memory. According to the energy output of 21 solar power plants, the results using deep learning algorithms show a superior forecasting performance compared to ANN. Kumar and Kalavathi 19 presented a proposed prediction model based on an ANN algorism and adaptive neuro-fuzzy inference system and used training based on historical data to validate the results of the proposed model. A real-time prediction model for the power of a grid-connected PV system was proposed in Su et al. 20 The proposed ratio model is able to fit the intermediate phase (9 a.m. to 4 p.m.) very well but is not accurate for the growth and decay phases. Hassanzadeh et al. 21 proposed a PV power generation prediction method using solar irradiance as the input, and the solar irradiance is modeled as the sum of a deterministic component and a Gaussian noise signal. The results show that using the spectral analysis method obtained good predicting results than those obtained with the ARMA model. Izgi et al. 22 presented a method of predicting the medium-and short-term solar power generation based on an ANN, which to determine the time horizon that with the highest representative for generated electricity prediction in different months.
This paper focuses on the following: (1) monitoring and performance evaluation of PV-power-generation parameters, (2) design of a set of online PV-power-generationparameter measurement and monitoring devices with simple structure, high sampling accuracy, small data fluctuations, and ease of measurement to monitor in realtime the electric parameters output in the operation process of PV power generation. The system is designed according to the structure of the PV power generation system, signal acquisition and processing, signal conversion and transmission, upper computer monitoring, and other aspects. A particle swarm optimization-based wavelet neural network (PSO-WNN) PV-power-generation prediction model is established based on the existing prediction methods, and the historical data of the power station is used to evaluate and predict the power generation. In the power generation process of a PV power generation system, its output power parameters shall be accurately monitored to meet the requirements of the power grid dispatching system and power generation performance evaluation. At present, common electric parameter signal sampling and measurement methods use shunt, Rogowski coil, and fluxgate. However, they have low measurement accuracy and relatively complex circuit structure, causing data fluctuations during sampling and measurement. By contrast, the new measurement structure of PV power generation parameters based on zero flux compensation principles can allow the accurate acquisition of electric parameter signals. The system comprises three parts: signal detection, signal processing, and compensating circuits. The sensors adopt an active electronic circuit compensation method in the compensating circuit to track the excitation current and compensate for the phase and amplitude to achieve a dynamic "zero flux" state. The sensors using zero-flux-compensation principles realize electric isolation between the output and main control circuit and overcome the measurement error of detection. The designed PV power generation parameter monitoring and experimental platform are shown in Figure 1. The monitoring system mainly comprises seven parts: PV cells, an inverter, a controller, output-end loads, signal measurement and processing module, a signal conversion as well as transmission module, and an upper computer.
The principle of the PV power generation parameter measurement and monitoring system is shown in Figure 2. First, the PV power generation parameter monitoring system mainly collects and monitors the output of DC electrical signals from the combiner box. Current sensors and voltage sensors based on zero flux principles and a signal processing circuit are encapsulated inside the signal measurement and processing module. Second, the sensors collect and transmit the output of DC electrical signals from the combiner box and the signal processing circuit filters and rectify the measured electric parameter signals to simulated output F I G U R E 1 Photovoltaic power generation parametermonitoring system and experimental platform. ① computer; ② 36w AC bulb; ③ 10w AC motor; ④ 28w AC motor; ⑤ 20.4w DC motor; ⑥ battery; ⑦ 50w AC bulb; ⑧ inverter; ⑨ controller; ⑩ DC signal measurement module; ⑪ AD Conversion module.
F I G U R E 2 Principle of the PV power generation parameter monitoring system. PV, photovoltaic. DC electric parameter signals. Finally, the signal conversion and transmission module are used to convert the collected analog electrical signals into digital electrical signals' output. The output terminal is connected to and controlled using an upper computer. The upper computer is used to receive PV-DC signals and perform analysis and storage.

| PV power generation parameter acquisition experiment
The PV power generation parameter acquisition experiment involves collecting the output of DC signals from the combiner box in the power generation process and transmitting them to the upper computer for storage and sorting. The PV power generation parameter monitoring system stores the collected current, voltage, and power data on the upper computer and establishes databases. In the PV power generation parameters monitoring system, the loads are connected to the output terminal of the inverter output AC electrical signals, which are directly measured using an oscilloscope. The loads are regularly shut down, and the power generation parameters are acquired. The shutdown sequence is as follows: turning on the small bulb, turning on small motor 2, turning on big motor 1, turning off the main switch, turning on the main switch, turning on big motor 2, turning off big motor 2, turning off big motor 1, and generating 20 sets of AC/DC power generation data. Comparison results between the DC parameter signals during PV power generation and the AC parameter signals output from the load end are shown in Figure 3.
By comparing the experimental results, we find the following: (1) all data items conform to the actual power generation situation, and (2) the actual output power of the AC end is significantly less than that of the DC end. The DC signal data graphics have small fluctuations and have no interference signals, such as harmonics. Therefore, applying sensors based on zero flux compensation principles to the PV power generation parameter monitoring system can realize real-time collection and storage of electrical signals in the power generation process.

| PV POWER GENERATION PREDICTION MODEL
3.1 | PV power generation mathematical model for engineering purposes PV cells are special diodes usually fabricated using semiconductor materials. A PV module is produced by connecting PV cells in series/parallel, tightly packaging them into an integral whole via a strict process flow, and installing the whole on one panel as a power generating unit as well as providing DC power for equipment. PV cells are influenced by sunlight intensity and panel temperature. Thus, the expression for the I-V output characteristics of PV cells can be written as follows: I f V S T = ( , , ). Considering the influence of temperature, the output I-V characteristic equation of PV cells is as follows: where I os is the dark saturation current and I ph is the photogenerated current; they are expressed as (2) and (3), where I SCR is the short circuit current of PV cells under standard temperature and sunlight intensity conditions, T is the background temperature of the PV cell panel, K I is the temperature coefficient, S is the sunlight intensity, E GO is the width of the energy gap of a silicon crystal, T r is the standard reference temperature (T r = 301.18 K), I or is the dark saturation current under T r , and A as well as B are ideal factors, generally A = B = 1.92. The five parameters included in the PV cell single diode equivalent circuit and its mathematical model (I ph , I o , R s , R sh , and A) are unknown and are related to temperature and sunlight intensity, R s , R sh are series and parallel resistors, respectively. These parameters are not included in the technical parameters provided by PV cell manufacturers and are practically difficult to determine. In practice, PV-cell manufacturers typically provide the five parameters (V oc , I sc , I m , V m , and P m ) under standard test conditions. Therefore, on the basis of the mathematical model for a PV cell single diode, the above technical parameters are combined with practicality and precision to establish a PV cell mathematical model for engineering purposes. This model simplifies the nonlinear mathematical model of PV cells and facilitates studies of the output characteristics of PV cells.
The parameters under open circuit and short circuit states are I = 0 as well as V V = oc and V = 0 as well as I I = sc , respectively. In the PV-cell-single-diodeequivalent-circuit model, R s « R sh ; thus, in practical cases, the serial and parallel resistances can be ignored. The parameters at the maximum power point are as follows: V V = m and I I = m . If the above conditions are substituted into the I-V output equation of PV cells, the simplified expression is given as follows: Typically, the range of sunlight intensity S is 0-1000 W/m 2 ; the temperature range of a PV cell panel is 10-70°C. When the sunlight intensity S (W/m 2 ) and the background temperature of PV cells T (°C) are not standard reference values, the relationship between the background temperature of the PV cells and the ambient temperature can be obtained from a large amount of experimental data via fitting. Therefore, the background temperature of PV cells under arbitrary sunlight intensity S and ambient temperature T ref can be given as follows: where K is determined by experimental data and the slope of T(S), usually taken as K = 0.03 (°C m /W 2 ). By reference to the technical parameters (V oc , I sc , I m , and V m ) of the PV cells under sunlight intensity and temperature, new mathematical model parameters (V′ oc , I′ sc , I′ m , and V′ m ) for arbitrary PV cell temperature and irradiance are obtained as follows: where e is the bottom number of the natural logarithm,

| Improvement of WNN
WNNs usually optimize network parameters with a single gradient, limiting the direction of parameter optimization and easily plunging network training to a local minimum. Furthermore, prediction based on WNN alone is less accurate, slow, and prone to network overfitting. In the three-layer topology of WNN, the output calculation formula for the jth node of the hidden layer is as follows 17 : where w ij is the link weight between the input layer to the hidden layer, h j is a wavelet basis function, b j is the shift factor of the wavelet basis function, and a j is the scalability factor of the wavelet basis function.
PSO is a heuristic intelligent optimization algorithm that has evolved from biotic group behavior. Each particle in PSO simulates one bird in the foraging of a bird group, which contains velocity as well as location information and represents one potential optimal solution to the optimal extremum problem. The velocity of a particle determines the moving direction and distance of the particle, which can be dynamically adjusted through personal and social learning methods to seek individual optimization in space and obtain the globally optimal solution. In a D dimensional space, we suppose altogether n particles constitute one swarm (X X X X = ( , , …, ) n 1 2 ) and the ith particle represents a vector in D dimensional space (X X X X = [ , , …, ] ) as well as the position of the particle in the D-dimensional space. Using the objective function, we can calculate the fitness value corresponding to the position of each particle X i . At this time, the velocity of , and the global extremum of the swarm is P g = [P g1 , P g2 , …, P gD ] T . During each iteration, particles constantly update their own velocity and position through individual extremum and global extremum. The update formulas are as follows: where w is the inertia factor, d D = 1, 2, …, , i n = 1, 2, …, , k is the number of iterations, V id is the velocity of the particle, c 1 and c 2 are learning factors, generally nonnegative constants, and r 1 as well as r 2 are random numbers distributed within [0, 1]. To prevent blind searching of the particles, the position and velocity of the particles are usually limited to [−X max , Adding momentum to the prediction network can further improve the network learning rate, accelerate the network convergence rate, and keep the algorithm stable. The formulas for the weight of the modified network with added momentum and the parameters of the wavelet basis function are presented below: (k is the momentum factor, usually taken as 0.95).
( ) , , , , PSO and adding momentum are combined to modify the weights of WNN and the parameters of the wavelet basis function. The learning process is as follows:

| Parameters and structural design of the PV-power generation prediction model
The output power of PV cells is mainly affected by external environmental conditions, such as solar irradiance and temperature. The expression of power generation per unit area of PV cells is as follows: where η is the photoelectric conversion efficiency of the PV cells under rated operating conditions, S a is the area of the PV cells, I is the total solar irradiance on the PV cells, ε is the temperature coefficient, and t is the ambient temperature. The impacts of different light intensities, ambient temperatures, relative humidity values, sunlight types, and seasonal changes on the output power of PV-power generation are analyzed as follows.
(1) Impact of sunlight intensity When there is sunlight during the day, the PV power generation system has positive output power. However, when there is no sunlight at night, the output power is almost zero. The relationship between PV power generation and sunlight intensity during the 12 h of sunlight within a day is shown in Figure 4.
As it is shown in Figure 4, the output power of the PV power generation system changes with irradiance. Compared with morning and afternoon, the light intensity is maximum at noon, and the PV generation also increases significantly. Therefore, the stronger the irradiance is, the higher the power generation is, and it is a positive correlation.
(2) Impact of temperature According to the historical database of the power station, the weather conditions for 7 consecutive days in April were sunny, and the other influencing factors had no visible changes, that is, the numeric values were similar. The average temperature and the corresponding output power of the 7 days are shown in Figure 5.
As it is shown in Figure 5, there is a large fluctuation range between the average maximum temperature and the average minimum temperature for the 7 days. As the temperature rises, the power generation gradually decreases, showing a negative correlation. This is because the ambient temperature increase causes a significant decrease in the photoelectric conversion efficiency of PV modules. Therefore, the output power of the PV power generation system decreases as the temperature increases when other influencing factors are constant.
(3) Impact of humidity A rainy day was selected for analysis from the historical database. Its sunlight intensity and temperature values remain stable, whereas the humidity value varies greatly throughout the day. The impact curve of the relative humidity on the output power is shown in Figure 6.
According to Figure 6, the relative humidity values at 7:00 a.m. and 6:00 p.m. are the highest and the changing trend of PV power generation is contrary to the changing trend of humidity. The increase in relative humidity causes a decrease in output power. This phenomenon occurs because the increase in relative humidity increases the water content in the air, and the refraction of sunlight in water vapor weakens the irradiance on the surfaces of PV modules. In addition, the thermal conductivity of PV modules is affected by relative humidity. The module cannot dissipate heat when there is an increase in humidity, causing an increased panel temperature that reduces power generation. Therefore, there is a negative correlation between PV power (4) Impact of sunlight Different sunlight types correspond to different quantities of air molecules, dust, and clouds in the atmosphere, as well as different irradiation areas on the solar panels. Thus, they have different impacts on the power generation of the system. From the 33 weather types released by the State Meteorological Administration, sunny, cloudy, and rainy days, which were the most frequent and the most common representatives within a year, were selected for study and analysis. According to the obtained data, the output power data of the same period in the same season with different sunlight types were selected for sorting and analysis. There is no sunlight at night when the PV-power-generation system is unable to perform normal power generation and has no power output. Thus, the power generation data for 12 h from 7:00 a.m. to 6:00 p.m. were chosen and analyzed. High noon is a period with the highest sunlight intensity and temperature within a day when the output power of the power station reaches the maximum. Figure 7 shows the PV power generation curves for sunny days, cloudy days, and rainy days.
As it is shown in Figure 7, sunlight is less blocked by clouds or particulate matter in the air on sunny days. Thus, the output power curve shows a relatively smooth parabola with small fluctuations and high output power as well as a high degree of similarity. On cloudy days, the sunlight intensity is reduced by clouds and other shelters. Thus, the PV-power generation curve shows large fluctuations and no longer follows a parabola and the output power is slightly lower than that of sunny days. On rainy days, the PV power generation (under the combined action of clouds, relative humidity, wind power, and other environmental factors) is lower than those of the other two weather types, and a weak regularity and a low similarity are shown. Therefore, predicting the output power of PV power generation systems on nonsunny days is difficult.

| Parameter determination for the PV power generation prediction model
Sunlight intensity, temperature, relative humidity, weather type, and season type have great impacts on the output power of the PV power generation system. To study the impact of adding irradiance to the input variables on the accuracy of the power generation prediction model under different algorithms, the input variables are divided into two types: with and without irradiance information. The topologies of the PV power generation prediction model with and without irradiance are shown in Figure 8. F I G U R E 6 Relationship curve between the daily relative humidity and the photovoltaic-power generation.
The number of nodes of each layer is determined as follows: the number of nodes of the input layer is the number of input variables of the model, that is, the number of dependent variables affecting power generation. The prediction model without irradiance uses 16 input variables, among which the first 12 input variables are a power generation time series from 7:00 to 18:00 of a similar day, and the last four input variables are the average temperature and relative humidity of the similar day and the predicted day. The prediction model with irradiance uses 26 input variables, among which the first 12 input variables are a power generation time series from 7:00 to 18:00 of a similar day, the middle 12 input variables are the sunlight intensity time series corresponding to the 12 h of the similar day, and the last two input variables are the average temperature of the similar day and the predicted day. The prediction model without irradiance information among the input variables finalizes the number of nodes in the hidden layer to 21, and the prediction model with irradiance information among the input variables finalizes the number of nodes in the hidden layer to 37. Because the prediction model is for predicting the power generation in 12 h of the next day, the number of nodes in the output layer is 12.
The performance of the prediction model must be evaluated overall after the completion of network training to extract prediction accuracy for numerical analysis. This paper presents an analysis of the accuracy of the prediction model using mean absolute percentage error (MAPE), relative error, and root mean square error (RMSE), respectively represented by the following formulas 2,10 : where N is the sample size, i is the serial number of data, P i f is the ith predicted value, and P i a is the ith actual value.

| Prediction results analysis without irradiance
The PV power generation prediction model optimizes the weight of WNN and the parameters of the wavelet basis  Table 1 were obtained. Using the parameters to train the network, the prediction results and the actual values generated good fitting outcomes. For the case of the data of sunny days input without irradiance among the input variables, the contrast curve and the relative error curve between the output data of the prediction submodels based on PSO-WNN and based on WNN and the actual data as well as the best particle fitness curve based on PSO in PSO-WNN are shown in Figures 9-11.
As it is shown in Figures 9-10, the predicted results of the output of the two models are close to the actual power generation trend. On sunny days, the relative error based on PSO-WNN neural network is less than the prediction result based on the wavelet neural network, and the maximum relative error based on PSO-WNN neural network is 20%, far less than the prediction method based on the wavelet neural network of 35%. Figure 11 shows that when the number of iterations is 30, the optimal learning rate is 4.178. Figures 12-15 show the  contrast curves between the output data and the actual power generation data of two prediction models without irradiance information among the input variables on cloudy days and rainy days, as well as the relative error curves of the two prediction models on cloudy days and rainy days.
It can be seen from Figures 12 and 13, on cloudy and rainy days, the difference between the prediction result based on the wavelet neural network and the actual value is larger than that based on PSO-WNN neural network. Figures 14 and 15 show that on cloudy days, the relative error of the prediction model based on the F I G U R E 9 Contrast curves between the predicted value and the actual value of two prediction models on sunny days. PSO, particle swarm optimization; WNN, wavelet neural network.
F I G U R E 10 Relative error curves of two prediction models on sunny days. PSO, particle swarm optimization; WNN, wavelet neural network.
F I G U R E 11 Best particle fitness iteration curve of the learning rate of PSO-WNN on sunny days F I G U R E 12 Contrast curves between the predicted value and the actual value of two prediction models on cloudy days. PSO, particle swarm optimization; WNN, wavelet neural network.
F I G U R E 13 Contrast curves between the predicted value and the actual value of two prediction models on rainy days. PSO, particle swarm optimization; WNN, wavelet neural network.
F I G U R E 14 Relative error curves of two prediction models on cloudy days. PSO, particle swarm optimization; WNN, wavelet neural network. wavelet neural network is much larger than that based on PSO-WNN neural network. On rainy days, the relative errors of the two prediction models fluctuated greatly, and the maximum relative errors increased compared with the relative errors on cloudy days. According to the above figures and the actual power generation data, we know: (1) On sunny days, the prediction model generates more accurate results than the other two weather types. This is because the sunlight intensity on sunny days is not affected by other factors, causing the output power trend of PV-power generation to exhibit a parabolic shape. Therefore, PV-power generation has a strong regularity on sunny days. (2) On cloudy days, cloud movement or an increase in cloud volume will affect the irradiance, leading to large fluctuations in the actual power generation values; thus, the correlation between the predicted values and the actual values is weak. (3) On rainy days, the model prediction results are the worst among the three types of prediction models because the uncertainty of environmental humidity, rainfall, sunlight intensity, temperature, and other factors will directly affect the PV-power generation. Therefore, there is no obvious correlation between the predicted value and the actual value, and the error range of the predicted values is relatively large when the interference of other factors is not fully considered.
The MAPE and RMSE of the PSO-WNN prediction model without irradiance are shown in Table 2 below.
According to Table 2, the prediction model has an error lower than that of the model based on WNN. The prediction accuracy of the proposed model is significantly improved over that of WNN, and the prediction results are closer to the actual power generation values. By comparing the prediction submodels of sunny days, the improved WNN has a MAPE value of 2.957%, which is lower than that of the original WNN. The prediction accuracy values of the other prediction submodels are also improved, and the relative errors at the morning and evening points are greatly reduced. Therefore, optimizing WNN with PSO can significantly improve the accuracy of the prediction model.

| Analysis of the prediction results with irradiance
After adding irradiance to the input variables, the contrast curve between predicted values and actual values, the relative error curve of WNN and PSO-WNN, and the best particle fitness iteration curve of the PSO-WNN-based prediction submodel of sunny days shown in Figures 16-18 is obtained. The evaluation results of the WNN and PSO-WNN-based prediction submodel of sunny days with/without irradiance among the input variables are shown in Table 3.
As can be seen from Figures 16-17, after increasing the irradiance, the predicted value based on WNN and PSO-WNN neural network prediction mode is in good agreement with the actual value, and the relative error F I G U R E 15 Relative error curves of two prediction models on rainy days. PSO, particle swarm optimization; WNN, wavelet neural network. based on the PSO-WNN prediction model is smaller compared to the WNN model. Figure 18 shows that after sunlight intensity was added to the input variables, the learning rate was optimal when the best particle fitness iteration number was 14. The iteration time was shortened, demonstrating that the convergence speed in the network prediction process was improved. Compared with the PSO-WNN model without irradiance among the input variables, the learning rate of the PSO-WNN network was improved and the relative error decreased, resulting in further improved prediction accuracy of the model. Therefore, comprehensively analyzing the prediction results of the two prediction models, it is found that adding sunlight intensity to the input variables is of great significance for improving the accuracy of the power generation prediction model. To verify the effectiveness of the proposed PSO-WNN prediction model in PV power prediction, the PV power generation and its error using the back propagation (BP) neural network, radial basis function (RBF) neural network, WNN neural network, and PSO-WNN neural network prediction methods are respectively compared under irradiation conditions, as shown in Figures 19 and  20. As can be shown from Figure 19, the prediction results of the PSO-WNN prediction model are the closest to the actual results, and its relative error is the smallest, compared to the other three prediction models. The predicted values of the BP neural network model and the RBF neural network model are different from the actual values, and their error fluctuations are also large. Therefore, the PSO-WNN-based prediction model is F I G U R E 17 Relative error curves of the WNN and PSO-WNN model on sunny days. PSO, particle swarm optimization; WNN, wavelet neural network.
F I G U R E 18 Best particle fitness iteration curve of the learning rate of particle swarm optimization-wavelet neural network. F I G U R E 19 Contrast curves of predicted value and actual value of different prediction models. BF, back propagation; PSO, particle swarm optimization; RBF, radial basis function; WNN, wavelet neural network.
F I G U R E 20 Relative error curves of different prediction models. BF, back propagation; PSO, particle swarm optimization; RBF, radial basis function; WNN, wavelet neural network. more effective in the field of PV power generation prediction.

| CONCLUSIONS
A PV power generation parameter monitoring system was established. Zero-flux sensors were applied to the collection and measurement of the combiner box DC parameters; this approach overcomes the measurement error of the sensor itself and improves the measurement accuracy. A PV cell mathematical model for engineering purposes was established; the WNN prediction algorithm was used to conduct short-term prediction of PV-power generation, and PSO and added momentum items were used to optimize the learning rate and weight of WNN. The impacts of sunlight intensity, ambient temperature, relative humidity, sunshine type, and season on PV output power were analyzed, and the outputs of prediction submodels based on PSO-WNN and based on WNN with and without irradiance were compared. The prediction model with irradiance was determined to have higher accuracy and smaller relative error than the model without irradiance. By comparing different prediction algorithms, the PSO-WNN-based prediction model was found to be more effective than the other prediction algorithms, with the predicted values closer to the actual power generation data. Through the research of PV power generation parameter monitoring systems and PV power generation performance evaluation, the functions of PV power generation parameter monitoring and performance evaluation are realized and provide complete solutions for maintaining the stable operation of the power grid.