Forecasting Solar Energy Production

Team members: Lee Shong Yan, Koh Si Yan, Chen Jianjun.

AI and Smart Nation

Energy is increasingly becoming a scarce resource. Hence, our solution involves using Artificial Intelligence/Deep Learning models to solve the efficiency of Singapore's overall power grid output. We use a Generative Adversarial Network (GAN) to predict solar energy power production in India's power plants. Optimizing our model provides us with the ability to forecast future solar power improving the management of power grid balancing, planning and optimization. Knowing the exact amount of power coming from solar also allows Singapore's grid to burn less natural gas for power and lower carbon emissions. Hence, the forecast of solar energy generation is vital for today's smart grid.

Overview

a. Exploratory Data Analysis (Dataset obtained from kaggle: https://www.kaggle.com/datasets/anikannal/solar-power-generation-data and Storm Glass API)

1. Data Preprocessing (i): constitutes preprocessing datetime variables in kaggle datasets and removing mismatching data between datasets
   Data Preprocessing (ii): extracting datasets from Storm Glass for the respective locations and dates and removing data that is in conflict with the kaggle dataset, as well as those that show anomalies

2. Data Visualization for both power and weather datasets

3. We use variables such as module temperature, ambient temperature, wind speed, cloud cover, humidity, and air pressure to predict irradiation and hence the AC power generated by solar cells

b. Our Model:

1. Benchmark done from a sequential neural network with irradiation as the ouput and the aforementioned variables as the inputs
2. The idea is to allow the model to implicitly learn complex relationships between the weather data and irradiation (and solar power) generated
3. Model is built in PyTorch and trained 

c. Accuracy of our model 1. The relative accuracy of the model shows a very complex relationship between the target and input variables, such that a very different type of AI model would be needed to sufficiently make predictions in this area.

Built With

• kaggle
• python
• stormglass