Improve methodology and clarify motivation

[camroach87]

In the methodology section, it is not clear at the outset whether you are seeking to do day-ahead predictions, 15min ahead predictions or build a descriptive model for scenario analysis. All three are hinted at and the notation is not carefully defined. In Table 2, you have temperature unlagged and lagged by 3, 6 12 hours but demand only lagged by days and week. Why not lag demand by 15mins if you want to do real-time forecasts and use smart demand-management? Presumably you do not want to do that, and it only becomes clear later. In fact it does not seem, on these early pages, like this is a predictive model, because the prediction point is not defined, but rather a reduced form structural model with lagged responses. You do not explain the context of the model sufficiently well, and the choice of lags and features appears to be casual.

Later it becomes clear that you are doing day ahead forecasts. It would have helped if that were stated at the beginning and the initial formulations thereby less vague. Why would a building manager want day ahead forecasts? With smart meters and smart controls I could imagine intelligent energy management systems being closer to real-time, particularly for trading in the Australian power market. It seems like this formulation is chosen without much thought.

• [x] Clarify what prediction I am doing (day ahead) in the methodology section. Check notation. Explain why day ahead forecasts are useful (demand response in commercial buildings, scheduling changes in set points, etc.).
• [x] Mention that hourly demand lags are not used as we are doing day ahead forecasts. Provide better explanation of why these lags and features have been chosen.
• [ ] Investigate "reduced form structural model" and update appropriately

[camroach87]

The problem of doing one day forecasts is mentioned in the second paragraph of the paper so it should be clear. However, it may be useful to mention this again in the model formulation section to avoid confusion.

[camroach87]

Response

The research problem of doing one day forecasts is mentioned in the second paragraph of the paper. However, it does help to mention this again in the 'Methodology' section while also explaining why it is useful in industry. I have added:

We focus on the problem of one day ahead electricity demand forecasting for commercial buildings. This has important applications such as allowing facility managers to adequately prepare a building for demand response (by adjusting set points and operational schedules) and reducing peak demand to avoid high capacity^[Peak demand events over a certain timeframe are often factored into a commercial building's electricity tariffs. These are known as capacity charges.] or time of use charges.

The choice of lags is now explained more in the methodology when introducing the pooled regression model:

When doing one day ahead forecasts we do not have have demand observations within the last 24 hours to use as lags and so we restrict our lagged demand variables to 1 day, 2 days and 3 days. The temperature variables can have 1-24 hour lags as temperature forecasts can be used to supplement the already observed data. Note that in this paper we restrict ourselves to using actual temperature data and not forecast temperature data to ensure our results are dependent on model formulations and not on any errors in temperature forecasts. In practice, forecast temperature data can be used instead.