Author: Suman Gautam
There is a rapid growth in solar energy generation, for example, according to EIA, State of Texas has planned to add 10 gigawatts (GW) of utility-scale solar capacity by the end of 2022. This means that there is also an increasing need to power management, both short-term and long-term. Managing power generation and distribution requires use of accurate forecast of energy generation across different sources. With increase in Solar power generation, forecast of solar energy generation would be critical to maintain power distribution and reduce short term power outage. A multi-scale solar irradiance forecasting may help to estimate the net solar energy generation.
This project is aimed to generate a predictive modelling algorithm that can forecast solar irradiance on short-term (hourly) to medium long term (10 - 30 days).
The for this project was acquired from Daily NOAA that hosts Surface Radiation data from 7 station around the United States 
.
The data are avilalble in the interval of ~17s of daily records. Although, there is yearly data going back few years, the primary focus of this focus is to predict on the short-term basis. Therefore, only data from year 2020-2021 was used for this project. The 2020 data was used as training data whereas the 2021 data was used for validation and test, since the 2021 data as of current is incomplete.
The size of data is huge was originally huge due to the recording at ~17s interval. Because we are interested in forecasting in hours and/or day to day basis, the data will be downsampled to hours only. The full data exploration notebook can be found in Data Exploration. Though, there are multitude of features available for this dataset, for this project only one feature that is associated with solar energy is used. In this case, 'Netsolar' radiation will be our time series upon which all the forecasting methods will be tested.
In this project, two different kinds of modelling approach were experimented:
- Conventioanl time series method such as ARIMA and SARIMAX:Modelling_I.
- Neural Network Architecture such as LSTM : Modelling_II.
Initial time series analysis indicated presence of seasonality in the data. However, SARIMAX method which accounts for seasonality doesn't seems to forecast as expected. This could be related to the lack of more historic data. On the otherhad, LSTM network seems to perform much better on this datata.
Time Series forecasting was performed on 24 hr, 10 days and 30 days interval. In general, the model tends to forecast better in the short term as can be seen in the result below:
Moreover, the forecasting used on hourly sampled data seems to capture the fine granularity of the trend very well.
For long term forecast, the model doesnt perform better beyond 10 days and the performace is satisfactory upto 10 days forecasting period:
Long term seasonality of solar irradiance is obvious, however, a short-term seasonality is also observed which was useful in prediction. Based on this, an hourly forecast can be made from this model to effectively plan and manage power generation and supply.
During the 2021 power outage in Texas, which was around February 7 to February 18 during which Solar Irradiance was also lowest. This pattern is cyclical and therefore a medium-long term forecast (upto 10 days) can benefit in monitoring the solar power generation. Potential alert to other source of power generation will help to minimize any weather related power disruptions, such as snowstorm.
Apart from global irradiance highs and lows, several consistent drop in the irradiance patterns have been noticed. It is suggested to collect past historic data to get more indepth insights significant drops. Especially, if they can be predicted.
Current dataset of 1 year period was originally used to predict short-term irradiance. However, the cyclical behavior of the solar irradiance can be useful in predicting long term and therefore more historic data is needed.
Current scope of work only allowed for univariate analysis, however there is a plenty of extra features that are available and are highly correlated to the solar irradiation. A multivariate method will be advantageous in creating accurate forecast especially several weather related variable such was wind speed, wind direction, cloud condition may directly affect the net radiation received.
├── data
├── images
├── models
├── Surfrad_data_collection.ipynb
├── Modelling_Sarimax.ipynb
├── LSTM_Modelling_All_Location.ipynb
├── Time_Series_Forecast_of_Solar_Irradiance.ppt
├── Miscellaneous_files
└── README.md