How does PowerPricePH work?
We input public data on daily electricity prices and supply & demand as covariates and get as output daily electricity price forecasts
Our models
We evaluated our models using the Mean Absolute Percentage Error (MAPE) but could not beat the 15% market error threshold during the time of the course. Some of the models, like the theta forecasting and naïve forecasting oversimplified the problem and didn't show any real trends. The volatility of the data was actually positive in these cases: the actual data oscillated around the predictions, so some of the predictions were actually pretty close. The Neural networks that we tested were an overcomplication to the problem at hand. The windows that it used to convolve the data did not map well to this problem set and the volatility of the data confused the layers.
Architecture
Based on public data from X to Y we trained different time series models. These models were saved on Hugging Face and the results imported in Amazon S3. Amazon Glue crawler was used to tabulate the data in AWS Athena. A JDBC driver allowed Tableau to link with Athena and query the data. A dashboard presenting the results was published on Tableau server and embedded into a website.
Key technical learnings
-
Volatility is very high for energy prices in Philippines and difficult to forecast, so we transitioned to a prediction window
-
Not all models are able to produce confidence intervals
-
Neural Networks were not the best approach for time-series forecasting
-
Covariates:
-
The only covariate we tested was Philippine holidays
-
Covariate data we considered are not publicly accessible: weather, stock markets, world energy prices (e.g. coal), exchange rates, supply and demand
-
Energy prices are highly volatile in Philippines
