The proposed methodology considers the following variables for defining the MLP input vector:
− time of forecast;
− recorded demand at current time;
− recorded demand at previous times;
− ambient temperature at current time;
− ambient temperature at previous times,
which will be explained as follows. The time of forecast is the time at which a demand value will be forecasted. It is represented by an integer number varying between 1 and 96, the value 1 corresponding to the interval between 0h00min and 0h15min and the value 96 corresponding to the interval between 23h45min and 0h00min (15-minute intervals used). The recorded demand at any time t (1 ≤ t ≤ 96) represents the average demand in kW registered in the interval associated with time t. The ambient temperature daily curve is treated exactly in the same way as the demand daily curve. The first three input variables listed above (time of forecast, recorded demand at current time and recorded demand at previous times) are mandatory, whereas the last two (ambient temperature at current time and ambient temperature at previous times) are optional. This allows for estimating the influence of ambient temperature on the load forecast quality, as will be seen in the next Section. Demand values for a particular MLP setup can be originated either from any of the substation transformers or from any of the primary feeders. Ambient temperature values are those recorded at the substation.
The output vector is defined using one or more values of future demand; that is, demand values at times t greater than the current time. Table 1 shows an example of an input vector and its associated output vector for:
− time of forecast t = 36;
− 1 current demand value (at t = 35);
− 3 previous demand values;
− 2 future demand values;
− no ambient temperature considered.
During normal operation, the MLP reads input vectors made available by the data-acquisition system and produces estimates of future load using the knowledge acquired during training (stored in its weights).