Short-term load forecasting method

Short-term load forecasting (STLF) predicts the load for the next hour to one week, mainly for unit commitment, economic dispatch and reserve scheduling. A widely used method is the time-series (similar-day / regression) method, and modern utilities also use artificial neural networks (ANN).

Multiple linear regression method (suitable and simple)

The hourly load is modelled as a function of the variables that affect it:

$$L(t) = a_0 + a_1 T(t) + a_2 L(t-24) + a_3 L(t-168) + a_4 D(t) + e(t)$$

where $T$ = forecast temperature, $L(t-24)$ = load at the same hour yesterday, $L(t-168)$ = load at the same hour last week, $D$ = day-type indicator (working day/holiday), $e$ = error.

Steps

1. Collect data: past hourly load (2–3 years), weather data, calendar (weekends, festivals like Dashain/Tihar).
2. Clean data: remove load-shedding periods and abnormal values.
3. Select variables: find those strongly correlated with load.
4. Estimate coefficients $a_0 \dots a_4$ by least squares.
5. Forecast using next day's weather forecast and day type.
6. Check accuracy using Mean Absolute Percentage Error (MAPE); good STLF gives about 1–3% error.
7. Update the model daily with new data.

Alternative

An ANN takes the same inputs (past loads, temperature, hour, day type), learns the nonlinear relation from history, and usually gives lower error than regression, especially for holidays and weather changes.