ENFOR

Electricity Load Prediction (LOADFOR)

LOADFOR (Load Forecasting) is used for predicting the load in power systems. As for the other ENFOR products LOADFOR delivers both highly competitive and accurate predictions, and unlike most other systems LOADFOR, together with the Quantile Regression Module, also delivers reliable estimates of the uncertainty of the predictions. Reliable estimates of the uncertainty is needed to obtain an economical optimal dispatch while keeping a very low risk of power failure. LOADFOR is optimal for short term and on-line predictions, with the horizon limited by the meteorological forecasts used.

LOADFOR is designed such that a joint setup and integration with WPPT (for wind power prediction), PRESS (heat load prediction), and PRICEFOR (price prediction) is optimal. Hence the uncertainty, time resolutions, input MET forecasts, adaptivity, horizon, etc. for all the products are consistent and the forecasts can be combined mathematically in scenario analysis and used as input for a global optimization.

History

At the Technical University of Denmark models for power load forecasting have been developed since around 1980.

LOADFOR is based on models developed for scenario analysis and long term forecasting. These models were developed at the Technical University of Denmark over the period 1994-1998. Subsequently, the models has been refined.

Characteristics

LOADFOR is built on artificial intelligence, and hence the system automatically calibrates to the actual situation. For instance the models automatically take into account changes in the user profiles, the the consumer behavior around holiday periods, changes in the use of cooling, changes in the MET forecasts, etc.

The models are typically a mixture of parametric, semi-parametric and non-parametric models. As an example the effect of holiday periods are most often described by using B-splines, which allow for a high level of flexibility around the start and the end of a holiday period. Another example is the effect of the solar radiation (direct and diffuse) which most often is appropriately described by a non-parametric function, since the turn off of light as a function of the radiative components can not adequately be described by a parametric function. Some of the functions in LOADFOR are static functions, whereas as other functions are dynamic functions. As an example the influence of the outdoor air temperature is most often dynamically due to heat accumulation in buildings - provided that a part of the power is used for heating or cooling. An example is included below.

Input to LOADFOR

Depending on the configuration of LOADFOR the following data is taking into account:

  • Online measurements of power load.
  • Met. forecasts of the local air temp., wind speed, solar radiation, etc.
  • Calendar information.

Output from LOADFOR

LOADFOR is tailored to the need of the end-users, and for the most often used configuration LOADFOR delivers:

  • Short term online predictions (say 0-120 hour ahead) of the power load.
  • Reliable confidence intervals for the future power load.

Elements of LOADFOR

This section is we shall briefly describe some of the most important elements, mostly models and methods, used in LOADFOR, and a couple of examples are considered in more detail.

Some of the most important models in LOADFOR is:

  • Models and methods for data validation

Since LOADFOR is used online, a rigorous checking of the input data is often needed.

  • Effect of holiday periods

For predictions of the power load the effect of holidays (summer holidays, Easter, Christmas, etc) is most often one of the most complicated tasks. Most often LOADFOR finds that the use of splines gives the most accurate predictions. Splines enables a high degree of flexibility by placing the knots around complicated time periods, but also by selecting the appropriate type of spline.

  • Day type effect

Clearly there is a difference between work days and non-working days, but most often we see a need for introducing an effect for each day in the week, since e.g. Monday is different from Tuesday, etc. A mixed effect between holiday periods and day type effect is also often needed, since the effect of the Christmas Eve depends on whether the Christmas Eve is located on a Sunday or a Wednesday.

  • Climate measurements / meteorological predictions

The climate variables taken into account most frequently are: Ambient air temperature, wind speed, and solar radiation. The models are in general non-linear, and in LOADFOR the dependency is most often described by non-parametric functions. Consider for instance the effect of outdoor air temperature (Ta) and the solar radiation R on the power consumption.

Figure 1 below shows contour plots of an estimated non-parametric function for the power load (in MWh /h) as a function of Ta and R.

Contour plot
Figure 1: Contour plot of the power load (in MWh/h) in Eastern Denmark as a function of ambient air temperature (Ta) and global radiation.

From figure 1 it is seen that a decrease in air temperature leads to a higher need for power in particular for temperatures below, say 12 deg. C. For the solar radiation it is seen that below, say 80 W/m2, we see a dramatic increase in needed power for decreasing lower solar radiation, whereas for a solar radiation above 80 W/m2, the power load does not depend on the radiation.

  • Thermal characteristics of the buildings etc.

The thermal mass of buildings implies that a rapid change of the outdoor air temperature does not lead to a similar rapid change of the needed power. This smoothing effect on the power load is seen with respect to both heating and cooling. Let us for instance consider the dynamic effect of air temperature and solar radiation estimated by LOADFOR. For a particular power system LOADFOR has estimated the step response functions shown in Figure 2.

Contour plot
Figure 2. Estimated step response functions on air temperature and solar radiation. The air temperature (Ta) dropped from 0 deg.C to -1 deg.C at t=0, and the global radiation (Rg) dropped from 25 W/m2 to 20 W/m2 at t=0.

The step responses estimated by LOADFOR in Figure 2 shows that for the particular power system a persistent change of the air temperature is not fully effective until 3 to 4 days after the changes occurred. For changes in the solar radiation the effect is seen to be fully effective after only 2-3 hours.

E-mail: info ¤ enfor . eu | Phone: (+45) 45 350 350