prediciton of residential PV Generation

Team members: Jingze Fu, Victor Santillan and Zhehong Wang

Clients: Dr.Truong X. Nghiem and Dr.Venkata Yaramasu

Mentors: Carlo R daCunha and Alexander Dahlmann

PV Work-breakdown structure (WBS)

Level 1

Photovoltaics

Level 2

Residential, Grid re-usage, and Prediction Software

Level 3

PV models, local power companies solar generation prices, machine learning, and programming

Prototype

In general, our case study and application models are used to accurately predict the generation of a provided PV system at hourly and 15-minute intervals, for the next 1-2 days, given the weather forecasts and other relevant information.

Griffin

Case study on non-data driven method of prediction of short-term PV generation and corresponding application of application model through Matlab.

David

Case study on data driven method of prediction of short-term PV generation and corresponding application of application model through Matlab. The current most suitable methods found are Wavelet-based Neural Networks with Genetic Algorithm optimization (GA-WNN) and Gaussian Process Regression (GPR).

Victor

Corresponding application of application models through Matlab on both parts of the case study, including non-data and data driven method.

The Project

We are to create a new method to predict pv generation so that households without battery storage don't have to rely on the grid. This can be scheduling major appliances like, washing machines and dishwashers during peak PV generation. The main problem, and the goal of this project, is to develop, implement, and validate data-driven methods (purely machine learning based or hybrid of machine learning and physics) for accurate, high-granularity predictions of residential PV generation.

Read More

Case study on non-data driven and non-machine learning

A method using weather information and regression analysis to predict the power generation of photovoltaic power systems in an energy network

non-data driven or non-machine learning.

based on weather information and regression analysis.

1.Weather: PV power generation varies with solar radiation, which in turn is strongly correlated with weather conditions. In addition, the conversion efficiency of solar batteries depends on the air temperature.

Solution :This method normalize the solar radiation (or generated power) by dividing it by the extraterrestrial solar radiation determined by location and time, and to use this as a solar radiation (output power) index. As a result, the attenuation of solar radiation reaching the earth’s surface is predicted, and weather data for other months can be used in forecasting

Region: The generated power of a photovoltaic system depends on the climate, terrain and other characteristics of the installation site. Hence, such regional dependence must be taken into account when forecasting the output power

Solution : this method described a time-series forecasting method using weather conditions as the explanatory variables

Different system: The power generated by a photovoltaic system varies with the orientation and tilt angle of solar panels, their capacity, and other design parameters. In addition, the out put power is affected by the conversion efficiency of the inverters.

Solution of Problem 2: This method estimated the solar radiation for every tilt angle, based on separation of the total radiation incident on a horizontal surface, treated as the objective variable in the analysis

Season: The power output of a photovoltaic system is affected by the seasonal variation of the orbit around the sun, for example, winter forecasts based on summer data would be inadequate

Solution: employed samples of past weather data for the site as well as current data (weather parameters and output power) collected during system operation. This method used precipitation, daylight hours as the weather parameters

Pros: minimize the influence of the outside world. It can be adapted to different weather, systems, seasons

Cons: 1.Both direct and indirect methods have unstable errors. In the regression analysis, the weather variables will affect the error of the analysis.

Regression analysis, as a more basic analysis method, has a considerable systematic error.

Case study on data-driven and machine learning

Recurrent Neural Networks (RNN)

How It Works

Cyclic constraints ensure that order information is captured in input data Predicts the PV power outputs in a peak zone only utilizing them observed in advance Prons

Share parameters in different time periods. Reduce training parameters and reduce computing costs.

gradient disappearance and explosion the input features used to estimate the PV power outputs in a peak zone could only consider a shorter time range from the time range of the prediction, making the problem more complex=> gate recurrent network (GRU)

Data Needed the information captured in the morning meteorological data : four meteorological factors comprise temperature, humidity, cloudiness, radiation. seasonal data : two seasonal factors are also included the month of the year and day of the month. PV power prediction in a peak zone using recurrent neural networks in the absence of future meteorological information – ScienceDirect

Gate Recurrent Network (GRU)

Recently, GRU has shown a successful, effective reduction of long-range sequence dependencies in many industrial domains

The proposed GRU-based model is designed To precisely capture both short-term and long-term dependencies between the PV power outputs in the peak zone and its preceding zone representing different time scales. To more effectively to catch the seasonal trends between PV power outputs in a peak zone and its preceding zone.

PV power prediction in a peak zone using recurrent neural networks in the absence of future meteorological information – ScienceDirect.

Support Vector Machine (SVM)

How It Works

expand

For linearly separable tasks, find a hyperplane with the largest spacing The main idea of SVM is to determine a function that has n deviation from the actual real vectors for training data and requires to be as smooth as possible. The key objectives structures of SVM are to minimize the training errors and reduce generalized errors to attain generalized performance. SVM is extensively employed to obtain good performance in classification as well as regression applications.

Prons: Kernel function can be used to solve nonlinear classification The classification idea is simple

Cons: the cubic SVM algorithm exhibited a bad performance, with RMSE of 21.72 and MAE of 15.667. SVM algorithm is difficult to implement for large-scale training samples Sensitive to missing data and selection of parameters and kernel functions Data Needed

solar PV panel temperature, ambient temperature, solar flux, time of the day as well as relative humidity, and the corresponding power of the solar PV panel. Solar photovoltaic power prediction using different machine learning methods - ScienceDirect

How It Works Gaussian process (GP) is a non-parametric algorithm, used for extremely nonlinear problems. It consists of random variables and accepts that all the input and output data have Gaussian distribution profile. GP provides distributions of all potential functions that are reliable with the training dataset. Therefore, the number of variables in a GP has no limits and increases with the number of training dataset. Prons 5/2 GPR algorithm provides a good performance with RMSE and MAE values of 7.967 and 5.302 respectively. For the uncertainty of prediction, the probability distribution of the prediction point value is directly output. Cons Squared exponential GPR exhibited poor performance, due to the complex relationship between the dielectric permittivity and the input parameters For the nonparametric model of Gaussian process, every inferential optimization requires matrix inversion of all training data points, which becomes impossible to solve when there is a large amount of data. Data Needed solar PV panel temperature, ambient temperature, solar flux, time of the day as well as relative humidity, and the corresponding power of the solar PV panel. Solar photovoltaic power prediction using different machine learning methods - ScienceDirect

LSTM - Long Short Term Memory

How It Works

Long Short-Term Memory (LSTM) networks are capable of computing an accurate approximation of the current state of observation given the past observations. Data Needed Average hourly dew point temperature , relative humidity , cloud cover , cloud cover, wind speed and east sea-level pressure were selected as weather variables

Literature review

Machine learning (ML) algorithms including support vector machine (SVM) and Gaussian process regression (GPR) were considered to predict the PV power based on input parameters including solar PV panel temperature, ambient temperature, solar flux, time of the day and relative humidity. To evaluate the performance of machine learning models under consideration, the root mean squared error (RMSE) and mean absolute error (MAE) were considered. Also, the feasibility of ML approaches were displayed by comparing the predicted data with real one. Cubic SVM, Quadratic SVM, and Linear SVM : large deviation and bad performance in RMSE and MAE Matern 5/2 GPR, Rational Quadratic GPR, and Squared Exponential GPR : small deviation and good performance in RMSE and MAE

Reference:
“Solar photovoltaic power prediction using different machine learning methods,” International Conference on Power and Energy Systems Engineering, 2021. Available:

https://www.sciencedirect.com/science/article/pii/S2352484721013287

Project Objectives

Objective 1: Develop Model for Acccurate PV Generation Prediction（Done） data driven and non data driven
Objective 2: Evaluate Model Performance (Done)
Objective 3: Provide User-Friendly Interface (Still can be optimized)

  • • The developed and implemented method(s) should be able to predict accurately the generation of a provided PV system at hourly and 15-minute intervals, for the next 1-2 days, given the weather forecasts and other relevant information.
  • • The achieved accuracy should be at the minimum: R^2 score at least 0.9 and RMSE less than 10% of the maximum PV output.
  • • The software implementation must be flexible, not specific to the said PV system, and can work with similar input data from any residential PV system.
  • • The software implementation must be easy to use with a friendly interface for entering specifications and input data, training models, validating models, and performing predictions.

Clean simulink mode for Non-data method

Table 1 Test Matrix for prediction using non-data driven model

Table 2 Test Matrix for prediction using data driven model

Table 3 Test Matrix for performance of software implementation design

Problem and solution：

Poster:

Countact Us