Jordi and Gerard from CIMNE-BEE Group spoke about, ‘Data driven models and time series analysis,’ looking at the accuracy of different models for the successful application of demand response (DR) services.

Spanning almost 30 years, CIMNE has gained reputation in more than 910 research and development projects and is dedicated to achieving the highest possible energy performance in as many buildings as possible. They aim to do this by the practical application of research results in a real business orientated environment.

Jordi and Gerard gave a brief introduction highlighting the different model types that are most commonly used in the field of simulation or modelling:

  • White box – uses physiological knowledge
  • Grey box – uses prior knowledge and information from data (white and black)
  • Black box – data-driven model based on inputs and outputs

CIMNE used a combination of the grey and black box model for the purpose of the Sim4Blocks (S4B) project.

The black box model is very complex and so to tackle this problem CIMNE used the statistical learning theory framework – used to gain knowledge, make predictions and decisions, or construct models from a set of data – which finds the most suitable model based on time-series data and estimates the uncertainty of each model.

The type of black box models included in the research were:

  • Static – linear models
  • Autoregressive – express the value of the next step based on previous time steps
  • Non-linear – identify real physical values/parameters of the building
  • Grey box – combination of physics and statistics

Examples of models for direct load control DR services

Direct load control are devices that can be operated by a utility or third-party energy provider to reduce a customer’s energy demand at certain times.

St Cugat, Spain, is one of three S4B demo sites and where CIMNE is looking to implement a test related to the heating and cooling of some offices in one building. The case study looks to optimise the consumption of the heat pump in line with the price of the electricity market using a grey box model which evaluates the heat transfer of the tank.

Results revealed that a linear piece-wise model was not good enough to capture the dynamics of this kind of data, for example the model could not calibrate accurately when there were temperature spikes, and so the project turned toward non-linear models. This model allows you to add as many parameters as you would prefer to capture the most accurate data.

Examples of models for indirectly incentivised DR services: flexibility in electricity consumption


  • Predict day-ahead load curve (chart illustrates the variation in demand/electrical load over a specific time)
  • Optimal cost scenario based on predicted load curve and forecasted electricity price
  • Estimate main domestic appliances’ usage


  • User behaviour models:
  • Phase 1: historical data from S4B users for representative daily load patterns
  • Phase 2: User time series classification from historical data
  • Phase 3: Prediction of daily load curve a day ahead
  • Forecasting – black box techniques – used incase the user behaviour model does not work as expected
  • Restrictions on some devices for example, users don’t tend to watch TV at three in the morning.

This service would potentially work for users through a device that would present a load curve predictive graph against cost, and an optimal graph against cost, to provide a comparison between the two to show users when the most cost-effective (cheaper tariffs) time to use a certain appliance is, for example, washing machine, oven, TV, chargers.