Testing data forecasting on different levels for granularity.
Forecasting is done using ARIMA (from com.workday:timeseries-forecast) and double exponential smoothing.
Methodology:
- Training data are retrieved from SparkWorks.
- Data are used split to a training part and a to-be-forecasted part.
- Data are forecasted using thge ARIMA model.
- The forecasted data are combined with the past data and double exponential smoothing is applied to the combined data array.
- The forecasted data are compared with the to-be-forecasted part to extract the sum-square-error.
Based on the tests the following parameters are considered the best for most cases:
- ArimaParams for model for data every 5 minutes: (p=?, d=?, q=?, P=?, D=?, Q=?, m=?);
- ArimaParams for model for data every 1 hour: (p=24, d=0, q=2, P=35, D=0, Q=2, m=7);
- double exponential smoothing parameters: alpha=1.0, beta=0.1
Legend:
- value: data to-be-forecasted
- forecast: result of the arima model
- forecast-smoothed: result of the double exponential smoothing for the forecast period
- Data points every 1h
- Forecast Period 336h (14 days)
- full data in SparksDrive/Test Data/ML Tests/TS-Forecasting
Examples:
- lp-temp-1
- lp-temp-2
- school-cur-1
- school-cur-2
- school-lum
- lp-water-pressure
