Forecasting energy consumption with no historical data when there is a trend

Question

I want to forecast new customers' energy consumption. Let's say I can construct a set of attributes to describe new and existing customers (e.g. size of business, type of business etc.) and I have the time series data on energy consumption of existing customers. However, I have no historical data on new customers' consumption.

If existing customers' consumption has seasonality but no trend, I could use the set of attributes to create a regression model to predict new customers' consumption based on existing customers' consumption. But what if existing customers' consumption has a trend? How do I know at which point of the timeseries I should start the forecasting?

Let's take the following example time series of an existing customer:

If I want to forecast new customers' consumption for one year, how do I know at which point of this time series to start the forecasting?

I do not have a dataset yet, so I cannot provide data examples. The answers could include an idea on what types of additional information I would need (if any) to solve this problem. Enrgy consumption is just an example, my question is theoretical and could be generalized to any other field. Also, names of python libraries to deal with this problem would be welcomed!

Thank you for the help.

Answer 1

Most time series models assume stationarity. Thus, you should account for trend and seasonality in your data and subtract the corresponding components from your original data. After that, you can model your normalized data also with arbitrary machine learning models; and forecast with new data the following months.

Carefully note that you may have features that explain the trend or seasonality component. If these components are also changing or expecting to change over time you could combine (e.g. add for additive time series models) predictions for all components. I.e. the sum of the trend, the seasonal and the residual prediction.

Importance of Normalization E.g. https://stats.stackexchange.com/q/31387/214243

Detrending Subtract the %yoy change for each month.

Seasonal decomposition E.g. http://www.statsmodels.org/stable/release/version0.6.html?highlight=seasonal#seasonal-decomposition

Forecasting There is an abundance of statistical time series models (AR, ARIMA, GARCH,...). But when data is normalized common machine learning models like decision tree or random forecast may also be applicable (see sklearn library for this)