Summary

pd4castr is a price forecasting web service that provides a price forecast by running a model formulated by a machine learning algorithm (MLA). A model is essentially a very big equation that a MLA derives by ‘training’ on historical data. pdView selects the input variables, the training periods and the type of MLA (as well as the hyper parameters used in the MLA) so that price forecasts can accurately predict prices for many energy market scenarios.

Input data can include forecast demand, available generation, interconnector flows and so forth. Input data may vary depending on the model.

The user interface of pd4castr is designed to provide a clear view and understanding of the price forecasts including model performance, input data and training periods.

Models are routinely retrained and released when there is a clear improvement in performance.

Forecast Time Periods

Currently only the predispatch time horizon is forecast.

Soon to be released is the week ahead forecast.

User Interface

Model re-training process

The pdView Team periodically generates new models by re-training a MLA. The re-training iteration can include any of the following:

• Introducing recent historical data,

• weighting historical time periods,

• changing or adding input variables, or

• changing the machine learning algorithm or the hyper parameters that tune the algorithm.

Open

Re-training the MLA

The key to retraining is to assess each new model against a known baseline. If a new model performs better than an old model then the new model is uploaded onto pd4castr where it can be used to generate new price forecasts.

Use Case

Early and accurate forecasts

pd4castr generates more accurate forecasts than AEMO’s predispatch price forecast when the forecast period is still some time away. This is because the pd4castr model factors in bidding behaviour and other dynamics whereas AEMO’s price forecast process follows a strict protocol and does not anticipate changes in unit commitment and bidding.

As real time approaches the forecast period, AEMO’s price forecast will likey become a better indicator of price outcome - this is because the dispatch period - which you can think of as forecast period time zero - is by definition (nearly) perfect since it is (nearly) always a 100% correct.

Independent forecast for middle and back office

Ideally risk metrics should be calculated independently of the front office. Where possible why not consume data provided by a reputable third party?

pd4castr is a price forecasting web service that uses a model that consumes input data to generate a price forecast.

A model is basically a very big equation that has been produced by a machine learning algorithm (MLA). Creating a model involves choosing a range of variables and then selecting historical periods (called training data) that are then inserted into the MLA. The training data include both input variables and the desired output - in this case price. The MLA then creates an equation that will most closely predict the desired output. There is a lot of time and skill involved in creating a new model, a MLA needs to be chosen (there are many algorithms available) that will accommodate the dynamics particular to the system being modelled. Then the MLA needs to be 'tuned' in a variety of ways (called hyper parameterisation). Finally the set of input variables needs to be selected that will capture the primary value drivers to predict the price.