By Chris Perullo, Turbine Logic

By Lea Boche and Bobby Noble, Electric Power Research Institute

Generation dispatch decisions are made every day with the goal of optimizing revenue from power while reducing operating costs. In simple terms, dispatch determines when to run generating assets and how much power each should produce.

In practice, determining how to dispatch generating assets to meet grid demand involves many disciplines. The generating capacity of the asset must be predicted, and the financial risk-reward equations must be balanced. Owners and operators express a degree of control over these two areas; however, external factors also exert influence.

The weather is a primary factor working against operators, especially in the summer months. Higher temperatures mean increased demand and reduced output from air-breathing assets, such as gas turbines. Accurate weather forecasts are also hard to come by, as anyone who has ventured out on a spring day can attest.

All of this leads to uncertainty in just how much power a generating asset can produce at a given cost. There is also asset-to-asset variation. As gas turbines age, they go through upgrade and maintenance cycles which means two gas turbines that were identical at commissioning may have measurably different performance 10 years later.

These uncertainties lead to assumptions in performance that can lead to reduced operating margins. Better quantification of variation in gas turbine performance, due to internal and external factors, will provide more precise information for dispatch decisions.

EPRI is working with gas turbine owners and operators to develop a software solution aimed at improving the ease and efficiency of making dispatch decisions. Through discussions with gas turbine operators, the team found two key areas where gross oversimplifications are often made: characterizing asset performance and estimating the impacts of variable weather on potential output.

The traditional method for predicting gas turbine performance uses correction curves. Most engineers have used these curves, which are often supplied at commissioning and provide estimates to performance vs. inlet temperature, pressure, and humidity.

The reality is that these curves, while presumably accurate at the first commercial operation date, are not updated often enough to accurately represent current performance. This can lead to misestimation of asset performance, which can have financial implications due to over- or under-committing to grid needs.

Figure 1 shows a comparison between correction curves from a gas turbine’s commissioning date and the current day performance. As the plant has aged its characteristics have shifted. Heat rate is underpredicted by the correction curve and the power output vs. inlet temperature has a different slope. While the changes are subtle, the ever-increasing uncertainty of renewables means those that can better quantify their gas turbine asset capabilities will be able to increase their margins in the marketplace.

To be fair, many operators do update performance curves but are often limited in resources and ability to process the noisy and incomplete data that is typical of real-world power plants. Many engineers also lack the training to properly curate, filter, and denoise performance data to extract a good performance model. EPRI and Turbine Logic have worked to leverage modern AI toolsets enabling an automated performance model.

Working with engineers at power plants led to the realization that most want to use AI, but do not want to be bogged down by IT and math. But the team has found a way to fuse physics-based models and AI to let operators perform a ‘click-one-button’ analysis that can accurately capture gas turbine simple and combined cycle performance.

Figure 2 shows the resulting accuracy of the model over an entire year compared to using static correction curves. Using a physics-informed AI model allows for good accuracy and can be used by non-experts in engineering and AI.

If your performance model is good, weather will have the largest impact on your predicted performance over the day. Being off by 10 degrees ambient temperature can mean a 5% change in potential power output. How accurate is your forecast? Have you ever actually thought about it? Or do you flip on the news every morning and take it at face value?

Finding reliable historical forecast accuracy information is challenging, if not downright impossible, so EPRI and Turbine Logic started tracking forecast accuracy in key locations. They found that the forecast accuracy varies in two repeatable ways. Error in the weather forecast was consistent with the wall-clock time of day and the number of days out from today.

Figure 3 shows the historical forecast accuracy for air temperature for Atlanta over a 3-month period. Obviously, the forecast is less accurate the further into the future you look, but it’s also less accurate from around 11 AM to 8 PM (20h), when afternoon thunderstorms can roll in. Predicting the time of arrival of storms is hard, but the drastic temperature swings can cause large variations in gas turbine output. Pressure and humidity follow similar trends. Better calculation of that +/- can help provide critical risk information to the groups responsible for dispatch commitments.

Even if you don’t have commercial dispatch optimization software, there are some basic things you can do now to better quantify the uncertainty in your performance predictions due to inaccurate weather forecasting. Quantifying error is a major first step and involves systematic, purposeful record keeping.

Typically, once a forecast is made, no one ever goes back and checks the accuracy and then uses that information to further improve the next forecast. EPRI and Turbine Logic have suggested some simple steps you can take to see where you stand. From there you can decide if you need to improve your modeling capabilities

Step 1 – Calculate your trade factors

If you have correction curves, figure out the sensitivity to inlet temperature and pressure. If you don’t have correction curves, filter your performance data to base load, and use a spreadsheet to fit a line. If properly filtered, your data should look similar to the data in Figure 2, but appropriate to your range of operations. If you want very rough estimates, you can use the following trade factors.

Table 1: Generic Gas Turbine Trade Factors

These trade factors are used to estimate a change in power output per change in operating conditions relative to the rated conditions of the gas turbine. For example, when your gas turbine is rated at 59 degrees F and 14.7 psia; you want to estimate the power output at 78 degrees and 14.4 psia using the generic factors above:

Change in temperature from rated = 78 – 59 = 19 deg F

Change in pressure from rated = 14.4 – 14.7 = -0.3

Estimated power output = (1 + -0.0035 x 19) x (1 + 0.0068 x -0.3) = 0.931

This means that you should expect to produce 93.1% of the power relative to the rated condition.

Step 2 – Record your weather forecast uncertainty

This should give you something to do during that morning coffee break. Every day, preferably at the same time of the day, start your weather app and record the forecasted temperature and pressure each hour for the next 24-168 hours (1-7 days). It’s your choice on how much to record. If you’ve got a clever intern or access to a forecasting API, you can also use those resources to pull the forecast every day. It’s best to structure the data in a spreadsheet as shown Table 2. Each day:

Table 2: Tracking Your Weather Forecast

Step 3 – Calculate the Forecast Uncertainty

Once you’ve collected a month or so of forecast data you can calculate the forecast uncertainty by the time of day or days ahead. If you’re using a spreadsheet program just calculate the standard deviation by day-ahead or time of day for the errors. It might look something like the below.

Step 4 – Calculate Impact on Performance (You can also skip right to this step!)

Once you’ve done the hard part of calculating your forecast error, you can plug it into the equation below to figure out the impact on power output. Or, if you think you know how bad your forecast is, you can just plug in different levels of forecast error to see the resulting impact.

So, for our example, on day 5 in the future and using our generic performance factors:

This means uncertainty in the forecast is leading to an error in predicted power output of about 2.8% This number may seem small, but considering a 500 MW combined cycle, this can be 15MW. Also, keep in mind that this represents the error that will occur less than 68% of the time. That means 1/3rd of the time your prediction will be even more inaccurate. Add on uncertainty in asset performance and there could be a significant mismatch between required and possible generation.

Evaluate the impact of your forecast errors on predicting asset capabilities and decide if it’s time to improve your processes.

Dr. Lea Boche is a senior technical leader at EPRI. She is a plant monitoring and diagnostics specialist and has authored a number of EPRI reports and conference papers on data science and digital twin applications.

Chris Perullo is the Director of Engineering for Turbine Logic. He leads day to day development of customized monitoring and diagnostic solutions and services for natural gas and renewable energy assets.

Bobby Noble is the gas turbine programs manager at EPRI and a Fellow of American Society of Mechanical Engineers. He is a key global leader in gas turbine diagnostics, and has authored a number of EPRI reports and conference papers on GT digital twin utilization case studies.