Skip to main content

Introduction

One of the main goals of creating Tensor Cloud was to accelerate renewable energy investment. Prerequisite for this is gaining trust and confidence of lenders and equity investors, which requires accurate, easy-to-understand, and standardized simulations. Creating simulations on Tensor Cloud can be done with a few clicks and calculations usually do not take more than a few minutes, even with hundreds of assets included.

You can use simulations internally, to better understand the performance of a portfolio of assets, or you can share them with external stakeholders when seeking investment into one or several assets.

What are simulations?​

Simulations on Tensor Cloud are long-term potential futures of the financial and technical performance of assets. The Tensor Cloud simulation engine combines all information you provide us about one or multiple assets, connected PPAs and scenarios to precisely simulate asset performance at 30-minute resolution over up to 40 years.

You can create simulations for multiple assets and across multiple scenarios at the same time. Simulations are always created in a way that they cover the entire lifetime of all included assets, from COD to decommission. Once a simulation has been created, you will not be able to edit its contents, except the simulation title.

Relationship with other building blocks​

As illustrated in the following chart, simulations on Tensor Cloud can contain an arbitrary number of assets, each of which are associated with one scenario and one PPA respectively. They are also the only building block that can be shared with external stakeholders without requiring creation of a user account on Tensor Cloud.

Methodology​

Price forward curves​

Tensor Cloud uses a combination of historical data and user assumptions to create long-term price forward curves. The process consists of three steps:

1. Baseline curve training​

First, we train a deep learning model to predict day-ahead electricity market prices for each grid area. For the model training, we use roughly 4 years of historical weather data (temperature, deviation from ideal temperature, GHI, wind speed) and JEPX day-ahead prices as inputs to determine the relationship between weather and market prices. Our training methodology automatically determines seasonality and auto-regression in the input data.

We are using weather data from the 5 locations in each grid area that have the highest predictive power for electricity market prices. These locations are determined based on our proprietary geospatial model of solar capacity concentration in Japan.

The resulting forecasting model is able to predict day-ahead prices based on weather data. This model is then used to generate the so-called baseline curves in the next step.

2. Baseline curve creation​

Our goal in this second step is to create baseline curves for each grid area. Baseline curves include the historical yearly, weekly and daily seasonality, but they do not include a trend, making them "neutral" or stationary.

First, we create weather data for the future. For this, we use historical weather data and project 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, this could be based on the weather data for April 1st, 2010. This is sometimes called time-series approach. We chose this method over using a typical meterological year (TMY) because it provides a higher statistical accuracy over a long time period.

We then feed the future weather data into the trained model from step 1 to derive baseline curves for each grid area.

3. Adapting curves to user assumptions​

Lastly, we apply assumptions made by our users in the Tensor Cloud scenario settings to adjust the baseline curves. Specifically, we use the average system price and average volatility assumptions to shape the baseline curves into the desired user-adjusted price forward curve.