Weather model

Background

Weather influences supply and demand for electricity. It determines the output of solar and wind powerplants and drives electricity consumption. Because prices are determined by supply and demand, weather therefore significantly influences prices on the electricity markets.

We have developed a unified, global weather model that creates realistic future weather data. When provided a location and a time range, the Tensor weather model creates weather data up to the year 2099.

Our weather model uses historical weather data and projects it into the future. That means, each point in time in the simulated future corresponds to a specific point in time in the past. For example, if we are creating future weather data for April 1st, 2030, then this day could be based on the weather data for April 1st, 2010.

This also means that our weather model is unable to provide meaningful short-term forecasts, as it is not a physical simulation of the planetary weather system, which would require a super-computer, nor a weather forecasting deep learning model like Google GraphCast.

As explained in the introduction, future weather data is used by the Tensor Cloud simulation engine as an input for

• Solar and battery storage asset simulation
• Creation of electricity price forward curves
• Curtailment simulation

Development goals

When we developed the Tensor weather model, our primary goal was bankability. For this, we focused on three main levers: explainability, statistical accuracy, and internal consistency. We believe that these are prerequisites for any given renewable energy or storage project to be bankable.

Explainability

Our weather model is not a black box. You can trace every year of future weather to a specific year in the past, by following our model's calculation logic below. We explicitly chose to not generate weather synthetically (e.g., by using a TMY approach) to enable data lineage tracing and increase your confidence in the Tensor Cloud simulation results.

Statistical accuracy

An additional benefit of our approach to generating future weather is that over a long time horizon, it is inherently more accurate than some alternatives that simply repeat the same year. The reason for this is that a continuous time series that repeats decades of past weather actuals captures the complete randomness of historical weather. That means that the digital twins of your assets could be exposed spikes of extreme heat or rain, similarly to how we observe them in reality.

There are trade-offs, however. While over long periods of time (i.e., years), our approach provides more accuracy, it can become problematic from an asset management perspective when the asset manager needs to compare real asset performance to simulation results on a monthly basis.

Here is where a TMY (typical meterological year) approach can be used to yield results that are more homogenous when looked at monthly resolution. Talk to us if you are interested in using TMY weather in addition to the Tensor weather model.

Internal consistency

When running simulations for

To simulate assets in the future, we need to provide weather data at any location in Japan in the future, and this data needs to be coherent with each other, so that different assets in the same zone experience similar weather, and so that all the assets experience the same prices with the correct weather correlation. To achieve this, we use weather data from the past and we re-index it to the future.

In our simulation, we use the weather patterns from the asset location to estimate solar generation, and we use weather from specific points to influence the price generation model, so that there is a correlation between solar generation at each plant and the prices that the plant will experience.

Example

Consider the case of a sunny day in Kyushu: the plant produces a lot of electricity, but because it's so sunny, many other plants are also producing a lot, and the prices go to zero. Without this correlation, there would be no way to know that the plant is not actually producing any revenue although it's at peak power output.

Climate change

Currently, the Tensor weather model does not consider the impact of climate change, as it merely repeats weather from the past. Our research team is looking into ways to incorporate long-term IPCC scenarios that include rising temperatures and more extreme weather events into our weather model.

That being said, Tensor Cloud is not a climate risk management platform, and given the degree of uncertainty around long-term climate predictions, we have opted to follow the current common practice in the renewable energy industry to simply ignore climate change in asset simulation models.

Calculation logic

Time periods in the past

For time periods in the past (for the case of assets with past COD dates until the present), we use past actual weather data from the asset's location.

Time periods in the future

For future dates, we generate weather by mapping past weather actuals those future dates. We take past weather for the same period of the year and we re-index it to fit the future dates. In particular, for each year $Y$ in the past we assign a new year $Y'$ defined as:

$$Y' = Y + (252 - \left \lfloor \frac{Y}{8} \right \rfloor) \cdot 16$$

This produces meaningful dates for any year until 2099 and ensures that closer weather dates in the past are used to simulate closer weather in the future.

info

Based on the mapping above, our platform maps future time to the following periods in the past:

_period 2024-2031 -> mapped to 2008-2015 _
period 2032-2039 -> mapped to 2000-2007
period 2040-2047 -> mapped to 1992-1999

We exclude any weather data from 2016 to the present because this data is used to train our models. This avoids using training data as input to the trained models.

Data sources

Our historical weather data is sourced from several datasets. For each location, we combine these datasets to get the most accurate historical weather conditions.

Data set	Spatial resolution	Temporal resolution	Data availability	Update frequency
ECMWF IFS	9 km	1 hour	2017 to present	Daily with 2 days delay
ERA5	0.25° (~25 km)	1 hour	1940 to present	Daily with 5 days delay
ERA5-Land	0.1° (~11 km)	1 hour	1950 to present	Daily with 5 days delay