BelVis PRO forecasts are based on an algorithm library which is co-influential in supplying today’s best available technology:
fuzzy and neuro-fuzzy
ANN (Artificial Neural Networks)
ALN (Adaptive Logic Networks)
ARMA, ARIMA, ARIMAX
similar day profile method
multivariate Linear Regression
Kalman filter
exponential smooting
The implementation quality of the methods leads to excellent forecast results at top computing speed. Some special Artificial Neural Network methods have been patented by KISTERS and so have potential for further performance enhancement. The main areas of application for BelVis PRO are:
administration of load profiles for individual customers, consumer groups, and total network loads
integral generation of forecasts, separated into balance areas, for use in portfolio and schedule management
generation of schedule profiles incorporating calendar-based constraints
power price forecast for EEX and OTC market products
creation of single customer or cluster forecasts
Neural analysis methods can also assist in the calculation of robust price forecasts for energy products at trading locations. Factors such as primary energy costs, available plant capacities, network constraints, energy demand, product-specific volume of trade, and specific events can be incorporated.
Calculation of forecasts
This can be done using the arithmetic basic functions. Load time series, also nested, can be selected from the stored collection as operands. Using the parallel forecasting option substitute forecasting methods can be configured. With this option various methods can be used to create and evaluate any number of forecasts.
Forecast and influencing factors
The influencing factors considered by BelVis PRO can be freely configured. The number of influencing factors in a forecast is generally not limited. Based on current practice, the effects of the following variables are included in making the sales forecast:
calendar-related criteria: day of metering (time of year), type of day (weekday), public/school holiday, summertime/wintertime, special events (regional peculiarities)
weather-related criteria: temperature (daily high and low), intensity of light or cloud cover, humidity, prevailing wind, wind speed, wind direction in any time resolution
customer group-related criteria: production schedules and ripple control programs
Load profile library
Forecasts can be made for new load-profile metered customers with no history of load metering on the basis of their customer group profile. Such group-based profiles are stored in the load-profile library as standardised profiles taken from representative load curves of various customers with a metering history, and from the effects of influencing parameters e.g. as a neural network.
Once history data starts to accumulate for these individual customers then this initial neural network is individually retrained step by step, factoring in the customer’s data, and thus gradually increasing the forecasting accuracy.
