NYISO Electricity Models using Tensorflow.js

This module is a collection of Tensorflow.js models that I've created in Julia using Flux but have exported to Tensorflow.js so that they can be used in browsers and via AWS Lambda functions.

The models included are:

• Electricity demand for each NYISO zone.
• Wind power generation (NYISO wide)
• Solar power generation (NYISO wide)
• Locational Based Marginal Prices (LBMP) for each NYISO zone

The models make forecasts at the ~1 minute, 5 minutes, 15 minutes and 1 hour ahead of time horizons. Each model produces 225 guesses about what the true value will be to form a probablity distribution.

Microprediction.org Prediction Bot

A simple Microprediction.org prediction robot is included in src/bot/bot.ts that submits the predictions from the models as part of the electicity prediction competition.

Examples

Utilize a model to generate predictions.

This example shows how to use a model to generate predictions:

import * as models from "@rustyconover/nyiso-electricity-models";
import moment from "moment";

async function exampleModelUsage() {
  // Load a model that predicts the overall electicity load/demand
  // for the entire state of New York an hour ahead.
  //
  // 12 forecast intervals ahead is an hour since each forecast
  // interval is five minutes.
  const model = models.getModel("electricity-load-nyiso-overall.json", 12);
  const target_time = moment.utc().format("YYYY-MM-DDTHH:mm:ss");

  // Obtain the weather information for the model.
  const weather = await models.getWeatherForModels([model], target_time);

  // Obtain the stream data for the model.
  const stream_data = await models.getStreamValuesForModels(
    [model],
    target_time
  );

  // Obtain the regressors of the model.
  const regressors = await model.regressors(
    target_time,
    weather,
    stream_values
  );

  // Using the regressors retrieve the predictions.
  const predicted_values = await model.predict(regressors);

  // The 225 predicted values form a non-parameteric probablity distribution
  // which express the model's prediction of the electricity demand
  // an hour from the current time.

  console.log(predicted_values);
}

exampleModelUsage().catch((e) => {
  console.error(e);
  process.exit(1);
});

Model Data Sources

The models use data from these sources:

1. High Resolution Rapid Refresh forecast products from NCEP/NOAA:

   • Temperature
   • Surface Pressure
   • 2 Meter Dewpoint Temperature
   • 2 Meter Relative Humidity
   • 10 Meter U/V Wind Components
   • Downward Short-Wave Radiation Flux
   • Visible Beam Downward Solar Flux
   • Visible Diffuse Downward Solar Flux
   • Total Cloud Cover
   • Low Cloud Cover
   • High Cloud Cover
   • Middle Cloud Cover

2. Existing electicity demand forecasts from NYISO.

The models where trained the continuous ranked probablity score used as the loss metric.

All of the code necessary to generate features, perform feature selection and train/test these models is open source.

These models are released under the MIT license and will be updated from time to time.

If you have feedback email [email protected].