Predicting Energy Consumption

Robust and transparent forecasting models are used to guide energy purchasing, sales and trade decisions

Industry Example: Electric Power industry, Electric power distribution, Smart metering

Who Cares: Purchasing, Sales, Finance, Electrical Utilitiies companies, Smart-metering solution provides, sectors where timely and robust long-term forecasting is required

What Enabled: Real-time long-term forecasting of electrical energy consumption scalable to the changing consumer population

Collaborators: CIOP Research Centre, (Cologne, Germany), GreenPocket GmbH (Cologne, Germany)

Showcase specifics

This case and the data was provided by GreenPocket GmbH and it's partner Spot Seven research group of the Cologne university of Applied Sciences. Green Pocket develops software solutions that process smart metering data to give consumers insights into their consumption habits and provide them with accurate forecasts of their future energy consumption. Among GreenPocket’s customers are electric utility companies who collect and analyze smart metering data of thousands of customers. This means that the energy consumption forecasting methods employed have to be scalable and efficient. The goal in the case study was to provide reliable and efficient long-term prediction of electrical energy consumption read from smart meters of a local community of German bakeries. Using only 12 weeks of quarter-hourly measurements the goal is to predict energy consumption 4 weeks, or 2688 samples ahead.

The data consisted of two variables - a time stamp taken at quarter hours from December 2010 to March 2011 (with missing data) and the recorded energy consumption in kWh. We used DataModeler Pro to build predictions and compare them with state-of-the art ensemble based forecasting methods. DataModeler Pro is extremely effective for creating transparent predictive models especially in cases with many variables of unknown significance. We used this unique feature to expand the original data set of two variables- time and consumption, into a wide data set with many time-lagged consumption variables - consumption N, N-1, etc. samples ahead.

To reliably predict energy consumption four weeks ahead we used time-lags of 28 days and earlier, i.e. time-shifts by 2688, 2689, 2690,..., 3367 samples. Because of the large backshift, we could only use 63% of the training data, but the reward was the models which could directly predict the response 2688 samples ahead without the need to calculate the prediction from the previous-sample predictions (and therefore accumulate possible prediction errors).

Here is an example of a predictive model for energy consumption E(t) generated by DataModeler Pro, assuming that time t is sampled over consecutive 15-minute intervals:
E(t)=8.94-253.62/(28.15+E(t-2688)+E(t-2689)+E(t-2690)+E(t-2785)+E(t-3360)).

Out of more than 670 candidate input variables DataModeler Pro has robustly selected only 10 drivers. Automatically created solution ensembles in the driving variables showed the best predictive performance in terms of all considered performance objectives when compared against an Automated Autoregressive Integrated Moving Average (ARIMA) method, exponential smoothing state space (automated ETS) method and the proprietary GreenPocket model.

The additional unique features of forecasting solutions generated by DataModeler Pro are transparency (the models capture an explicit underlying dependency of previous lags to the consumption of interest), efficiency (explicit formulae are instantaneously computable providing real-time prediction), trustability - all prediction come with quantified confidence intervals, which allows early warnings, scalability - as the community of energy consumers grows, the developed model structures allow real-time optimization to the new consumption patterns.

Related Risks, Publications, Resources

Key Words: nonlinear time-series forecasting, explicit forecasting models, energy, smart meters

Highlighted DataModeler/ DataSolutions Features/ Capabilities: Modeling non-linear heteroschedastic time series, robust variable selection, transparent prediction models, prediction many samples ahead, ensemble methods, missing data handling