Access Official Technical PDF

Share Summary:LinkedInEmail

Accurate mid-term load forecasting helps utilities plan generation and operate power systems around expected demand. For residential feeders, that task is complicated by changing weather, appliance use, household routines and seasonal patterns. A case study of a residential sector in Mosul, Iraq, examined how an artificial neural network could forecast daily peak load one week ahead using historical demand, weather information and a more detailed representation of household consumption.

Why residential load forecasting is challenging

Electricity demand does not follow a simple, fixed relationship with time. Temperature can shift the use of cooling equipment in summer and electric space heaters in winter. Daily routines and the appliances present in each household also affect consumption. As a result, a forecast based only on recent total load may miss important changes in the factors driving demand.

Forecasting horizons serve different operational needs. Short-term forecasts typically cover hours to a week, mid-term forecasts extend from about a week to several months, and long-term forecasts look further ahead. The Mosul study focused on a one-week-ahead forecast of average daily peak load for a residential feeder. Such estimates can inform power-system operation and help utilities anticipate demand on the distribution network.

Using a neural network for mid-term load forecasting

The forecasting model used a feed-forward, multi-layer perceptron (MLP) trained with supervised backpropagation. An MLP contains input, hidden and output layers. During training, the network adjusts its weights and biases to reduce the difference between its forecast and the observed load. This approach is suited to problems where the relationship between demand and its influencing factors is nonlinear.

The study used one year of data, collected from April 2010 to March 2011 for a selected residential sector in Mosul. Information included feeder load, household appliance use and daily maximum and minimum temperatures. Before training, erroneous load points were removed using a moving-average method, and the data were normalized for use by the network.

Why disaggregating residential load matters

A key feature of the model was its treatment of demand as more than a single feeder total. The daily peak load was divided into five components:

  • Lighting: indoor and outdoor lighting.
  • Domestic use: everyday appliances such as refrigerators, kitchen equipment, washing machines, pumps, televisions and computers.
  • Cooling: air conditioners, coolers and fans.
  • Space heating: electric heaters used primarily during cold weather.
  • Water heating: water heaters, treated separately because they are used across seasons.

These categories reflect how residential demand changes over the year. In the study’s selected months, cooling represented a large share of summer consumption, while space heating and water heating contributed substantially in winter. Separating these uses gives the model information about the composition of demand, rather than asking it to infer every seasonal shift from the total load alone.

Inputs, training and evaluation

The model combined load and calendar information with weather variables. Its inputs included the current day’s feeder peak load, the five load-component shares, the day of the week, the sequential day number, and maximum and minimum temperatures for both the current day and the forecast week. The target was daily peak load for the following week.

Of 351 data points, 60% were selected randomly for training and the remaining 40% were used for validation. The researchers tested different hidden-layer sizes, transfer functions and training algorithms. Their selected network had 12 input nodes, five hidden neurons and one output, with sigmoid functions in the hidden layer and a linear function at the output.

Forecast performance was evaluated using mean square error (MSE), mean absolute percentage error (MAPE) and correlation between predicted and actual load. These measures describe different aspects of model behavior: MAPE expresses average error relative to actual demand, MSE penalizes larger deviations more heavily, and correlation indicates how closely the predicted and observed values vary together.

Results and practical interpretation

For the reported one-week forecast, the model achieved a MAPE of 1.921%. The paper also reported MAPE values of 3.39% for training and 4.58% for testing, with a testing correlation coefficient of 0.891. These results demonstrate the potential of the proposed input strategy for the specific residential sector and dataset studied.

The figures should be interpreted in context. The model was developed using one year of data from a particular feeder, so the reported accuracy does not establish the same performance for other locations, customer groups or operating conditions. Forecast quality depends on the input data, model design and characteristics of the load being predicted. The study itself notes that residential demand was variable and non-stationary during the observation period.

Implications for utility and automation teams

For engineers and system planners, the case study highlights several practical considerations when developing a feeder-level forecast:

  • Represent the drivers of demand. Appliance categories can reveal seasonal changes that are obscured in aggregate load data.
  • Include relevant weather and calendar information. Temperature and day-of-week indicators help describe conditions associated with different demand patterns.
  • Prepare data before training. Outlier handling and normalization are part of building a usable forecasting model, not optional finishing steps.
  • Validate on data not used for training. Separate validation data provides a check on how the model performs beyond the examples it learned from.
  • Use multiple evaluation measures. Error percentages, squared error and correlation together provide a more informative assessment than a single score.

The study also suggested that combining the neural-network approach with wavelet transforms could be explored in future work. This was presented as a possible avenue for improvement, not as a tested result of the reported model.

Conclusion

The Mosul case study shows how an MLP can support mid-term load forecasting when its inputs reflect the structure of residential demand. Decomposing feeder load into lighting, domestic, cooling, space-heating and water-heating components, then combining these with temperature and calendar data, produced a one-week-ahead forecast with a reported MAPE of 1.921% for the study’s forecast period. For utility and automation teams, the practical lesson is to align model inputs with real consumption drivers, prepare and validate data carefully, and treat reported accuracy as specific to the feeder and period on which a model was developed.

Access Official Technical PDF

About The Author