Combining a Weather Forecast with a Stochastic Weather Generator

Posted by Jason Lillywhite

A couple of years ago, I blogged about the application of the GoldSim WGEN model for producing probabilistic time histories of precipitation and temperature data. WGEN is useful for modeling natural systems impacted by uncertain climate influences. This model can also be used for simulating future scenarios and forecasting.

Even though the results of a WGEN model are probabilistic and uncertain, they are based on historical observed records and tied to the latitude of the earth. This allows us to produce realistic future simulations. But if you plan to use WGEN for short-term forecasting (less than 6 months), it might be helpful to incorporate a weather forecast for the first part of the simulation. Read more to see how a web service forecast model is combined with WGEN to produce a more reasonable forecast for a specific location without having to manually download the forecast.

To develop this model, you will need to combine 2 existing models from our library:

• WGEN Weather Simulator
• Weather Forecast from Web Services

The WGEN GoldSim model is a daily time step model that can be run for long-term simulations. The GoldSim implementation is currently built and tested for 1-day time steps. Because of this, the forecast model was also set up for daily time steps. If you need to use something different, you will need to further modify the above models.

The forecast model provides daily min/max temperatures along with the chance of rain for 7 days. In addition to this, it provides the forecasted rain amount for 3 days. The temperature forecast is applied to stochastic elements that use the mean absolute error (MAE) values published by the National Weather Service to define a triangular probability distribution where the most likely value is the forecast then the min and max are the forecast +/- the MAE. Below is a screen capture of the forecast model by itself:

The next step is to combine WGEN and the forecast model into a single model that transitions from the forecast to fully probabilistic weather simulation after the first 7 days. Here is a screen capture showing a new model with the 2 above models pasted in.

You can connect these 2 models by applying conditions on the results that switches from the forecast result after it is complete then switch over to the WGEN results. An example Selector is shown below that provides this type of switch for the rain generator input, probability of rain ("Prob_Wet"), that switches from the NDFD forecast rain probability (from NWS Web Services) then immediately switches over to the WGEN output.

This will prove to be quite an abrupt jump for some simulations as you transition from the forecast to the probabilistic weather generator. 

Update (15 August, 2020):

I implemented a transition period in the model that provides a linear transition from the last forecast value to the WGEN output over a specified transition duration. The image below shows the result of this transition for a single realization. On the last day of the forecast, there is 100% weight factor on the forecast and 0% weight on WGEN. At the end of the transition period, it is the opposite: 0% factor on the last forecast value and 100% applied to the WGEN output. Through the transition period, it changes linearly over time.

Below is a screen capture of a plot for just 1 realization so you can see how the transition occurs during the time from the last forecast value until it merges to the WGEN output. 

You can adjust the duration of the transition period using a slider on the dashboard.

The result of this is a more gradual transition from a published forecast to a stochastic output. In my case, I added containers with the word "transition" to hold the functions that output the values as the simulation time transitions from the forecast period into the future stochastic period. 

The result is a model that starts with a deterministic output from the weather forecast then transitions to probabilistic as demonstrated in the dashboard of this model, shown below. This type of simulation is more accurate for short-term forecasts than using WGEN alone because it's taking into consideration actual forecasts.

If you would like a copy of this model, please let us know in a comment below.