This repository contains an ensemble of machine learning models designed to classify ideal locations for installing utility-scale solar power plants in the state of California. This project represents our submission for the 2023 Erdös Institute Data Science Bootcamp Final Project.
AUTHORS: Utkarsh Agrawal, Mansi Bezbaruah, Harsha Hampapura, Akshay Khadse, Meysam Motaharfar
This project is an innovative solution designed to streamline the site selection process for utility-scale solar energy farms. Leveraging geospatial data analytics and machine learning algorithms, our predictive models assess various factors, including: modeled average annual electricity output, topography, land use, and population. Then, we provide a recommendation system for potential solar power farms, on which our stakeholders can conduct a few extra studies before investment and construction. This system significantly reduces time and resources spent on research, and thus, is an indispensable tool for our stakeholders.
Key Performance Indicator: How efficiently does the model predict suitable locations compared to actual installations?
Our data consists of 5 features, namely- modeled average electricity output, land cover data, elevation, slope and distance to a nearest city. The solar output tells us about the theoretical maximum amount of solar energy that we can harness at this particular location. The elevation and slope are indicators of how difficult or easy it would be to build a solar farm. The land cover informs us about potential land use restrictions at this place, example, open water, or dense forest. Finally, the distance to the nearest city serves as a proxy for electricity demand, labor and transportation cost in building the solar farm.
For plots any of our features on a map please visit figures.
Cleaning:
To train our classifier model, we also need information about places where solar farms cannot exist. To this extent, we set thresholds for each feature such that beyond these threshold values, it would not be practical to build a utility-scale solar farm. This helped us identify various locations across California where solar farms cannot exist. We also collected data on a feature called "ac_annual_ouput". The modeled AC output indicates the maximum solar energy harnessable at a given location for a 1MW capacity farm, factoring in typical losses and weather conditions, sourced from the National Renewable Energy Laboratory's solar model. This approach negates the need for separate training on variables like solar irradiance, temperature, and cloud cover. Using the annual average accounts for variations in solar output over shorter periods.
Our data collection and cleaning code is present in new_data and data_code. Moreover, all data along with API calls are present in data.
Once we had our data cleaned and preprocessed, we plotted density functions for all our features and then plotted the features against the ac_output variable. These graphs are present here. We observed that the features are not Gaussian and don’t show a clear linear or binomial correlation. This helped us eliminate certain models. We searched for models that are suitable for a classification task. We chose Logistic Regression, Decision Trees, Support Vector Machines, XGBoost Classifiers, and Deep Learning with classification layers .
Modelling code, detailed metrics, and confusion matrices are present in ml_code. We also provide a detailed report here.
As stated above, our goal was to create a model that can classify if a given location is suitable for a utility-scale solar farm. After carefully collecting and cleaning the data we were able to train models that classify with 95% for the existing solar farms. Another observation from our project is the importance of various geographic and climatic conditions in building a good solar farm. This can help a shareholder to make decisions in case there is a dilemma over building a solar farm based on a couple of contradictory features. Surprisingly, we find that all features contribute equally to a location being fit for a solar farm.
There are various ways we can further improve our model.
- Adding Features: There are other features that we would like to include in the model but we were unable to do so for various reasons. We could include commercial aspects such as, land cost, human labor and transportation cost, profit generated, etc into the model. Moreover, we would like to quantify measures like government support, for example, by introducing a metric and checklist for local laws and incentives towards solar energy.
- Including more locations: Due to time constraints, we had to restrict our data collection to the state of California. We could further train and test our model on locations outside california. This would increase the accuracy of our models and would provide us with more points to test the efficiency index on.