Short-term memory

The Hybrid Multi-Layer Radial Basis Function neural network (HML-RBFNN) is designed exclusively by the Sideratos Forecasting company.

It is an adaptive, self-constructive neural network with an on-line training algorithm.

It does not require any historical data for training. It learns very quickly and it has good performance after the first two weeks.

It retains the abilities of the RBFNN and ARTMAP networks, while it solves their drawbacks, like the curse of dimensionality or the generalization deficiency.

The HML-RBFNN has four layers: two non-linear layers with radial basis functions and two recursive linear layers.

The first layer maps the sections of the artificial brain. It makes a data clustering depending on the main attributes of the problem: For example, in the wind power forecasting case, the main attribute is the wind speed or in the load forecasting case the main attributes are the calendar elements.

At the first layer, more than one neuron is activated in order to save the same information in different sections of the artificial brain. This is a very important functionality of the human brain. The information is firstly preprocessed and then is saved in the second layer where the sections are located.

During the second layer training, a genetic algorithm tunes the neurons’ radius (each neuron’s element has a different scale) and the neurons’ vigilance (which represents the neuron’s membrane potential). Also, the Orthogonal Least Squared algorithm is applied for the generalization check.

Every section of the second layer has a linear output which is located at the third layer. At the fourth layer all the sections outputs are combined giving the requested prediction.