Accurate load forecasting is a high priority for Transportation Companies, Shippers and Distribution Companies. Expectations regarding attainable accuracy often exceed what is possible given the data quality, data availability and other drivers about which we know very little. Why is this? This paper tries to answer this question by describing the forecasting process and highlighting those things that can affect the forecast accuracy. The paper also describes how a user might get around some of these affects. Some examples of real data are described together with what forecasting performance we might reasonably expect to get using such data. Some examples are discussed where it is unlikely without direct communication with the customer that a better forecast can be produced than just using yesterdays value. It is important to understand that accurate forecasts cannot be obtained easily by throwing data at a forecasting package. The conclusion is that time must be spent understanding the problem with end-user input before forecasting models are built. Having had this dialogue with the end-user, reasonable expectations can be set and a forecasting system delivered that fully achieves them.
This paper explains the pitfalls that can be encountered when attempting to produce accurate demand forecasts. Forecasting is a relatively simple process, it requires you to build a model relating an input, maybe historical load, to an output possibly future load. What can be more simple? The problem is that this is not the whole story. To explain further the forecasting process is described in a little more detail, followed by some examples of where forecasting is easy or not so easy and finally a few remarks are made on whether accurate forecasting is a reality or just a dream.
Forecasting can be divided up into six clear processes:
analysing data and identifying models
estimating parameters or training models
verification of models
If any of these processes is not undertaken with care and a certain amount of skill then there is potential for an inaccurate forecast. Each process will be taken in turn and the potential pitfalls identified.
Obtaining data sounds like the most straightforward of the tasks. In reality, this is not so. There are many things to consider before we can be certain that you have all of the data that you need and that the data is of the necessary quality. In order to forecast any load some historical data will be required, for some loads a few days of data will be sufficient to forecast the future. An example of this might be a constant process load where demand tomorrow will be the same as the demand yesterday or perhaps the same day last week. For others the load might be weather sensitive so we may need to include historical weather data in the model.