USING CHAOSHUNTER TECHNICAL INDICATORS FOR LOAD FORECASTING
If you are struggling with a forecasting problem that involves time series data, ChaosHunter can produce a readable formula that will impress your boss. You simply use technical indicators from ChaosHunter to find relationships in your data to generate accurate predictions.
It is not unusual to select the technical indicators in ChaosHunter as potential inputs to trading models. However you can use some of those same technical indicators to build load forecasting models for processing operations such as a water treatment plant or a manufacturing facility if you’re working with time series data.
We had an inquiry from a power company that had several years’ worth of data which they wanted to use to predict power requirements one hour ahead of time. The prediction model had to be a readable formula that could be used in a spreadsheet. The company had collected data on hourly temperatures and meter readings, day of the week, season, and the time of day. They wanted to improve the accuracy of their forecasts.
We selected ChaosHunter to build the prediction because of the requirement for a readable formula. When selecting functions for possible inclusion in the formula, we began with the arithmetic and algebraic categories When looking for potential independent variables for the model, we decided to experiment with some of the technical indicators to capture relationships in the data from one period to the next. We included the Momentum (change), %change, spread%, regression slope and lag technical indicators that were applied to the temperature and current meter reading values in the data. ChaosHunter will decide how many periods are used in the calculations.
With the data file open in Excel, we added another column that was the lead 1 hour ahead of the current meter reading. We copied the current meter reading column and pasted it to the next column, shifting the data up one row. The result is that each data row now includes input values for the current period and the output we’re predicting is the meter reading for the next hour. Each row in the data file becomes a set of historical patterns between the inputs and predicted load which ChaosHunter can use to develop a formula for the relationship.
We used an optimization goal function of minimize mean-squared error so the predictions would match the actual meter readings 1 hour ahead of time as closely as possible. We selected the Evolution Strategy genetic optimization method and a population size of 300 to allow many different solutions to be tested.
ChaosHunter developed the following formula which achieved an R-squared value of 0.994993:
The formula included the momentum (change) of the meter reading over 12 periods and the slope of the linear regression line of the meter reading over 3 periods.
The actual meter readings 1 hour ahead of time and the model’s predictions were in close alignment.
