Before defining what spatio-temporal forecasting is, it is best to define what spatio-temporal means and what forecasting is.

Spatio-temporal

Basically, spatial refers to space and temporal refers to time. In data science, spatial temporal, or spatio-temporal term is used when the data has both space-related properties (the spatial dimension) and time-related properties (the time dimension). As an example, energy consumption of a city can be classified as spatio-temportal. Such data is expected to have time dependent behavior such as monthly/daily/hourly at different districts of the city. Hence, each row of this data will correspond to a certain time-frame (the time dimension) and to a specific region of the city (the space dimension), yielding a spatio-temporal structure. For example, one row of this data may indicate the amount of energy consumed in the Midtown area of Manhattan, NYC on the 15th day of December 2021. Granularity for space and time may be higher or lower depending on the data set.

Forecasting

In dictionary, forecasting is defined as a statement about what someone thinks will happen in the future. In data science, forecasting is defined as estimating the future events based on the past and present data, most commonly, by analyzing the trends and fitting a formula that gets close to the data as much as possible.

Combining the two definitions, spatio-temporal forecasting is about estimating the type of event that will occur at a specific time and location in the future. For the energy consumption example, estimating how much energy will be consumed tomorrow in the Midtown area of Manhattan, NYC is considered a spatio-temporal forecasting.

For producing spatio-temporal forecasts using data science approaches, past data need to be analyzed and appropriate mathematical models should to be generated, which can define past events as a function of spatial and temporal properties of the data. For example, if we try to predict the energy consumption at a specific district for the next weekend, then we need to find a mathematical model that outputs the amount of energy consumed as a function of time-of-day, day-of-month, month-of-year, temperature, demographic information, economic indicators for that district etc.

Finding the mathematical model that will produce the most accurate results for a spatio-temporal forecasting problem can be a simple task in certain cases, but in general this is a very challenging goal to achieve. Especially for cases where expert knowledge exists and can be used to produce reasonable forecasts (for example the expert at an energy provider company can predict with a reasonable deviation how much energy will be consumed tomorrow without using any mathematical formula), custom deep networks are required to find the best mathematical model that will surpass the forecasts of the expert. Based on all our expertise in this field, we believe that this is art and requires artistic mastery besides high quality engineering.