Advanced analytics forecasting for alternative energy

Supplying and distributing alternative energy — whether from wind, solar or hydropower — is a complex business. Too complex, in fact, for today’s commercial offthe-shelf software (COTS) solutions.

Some of this complexity results from the inability of alternative energy suppliers to gain a future view on energy supplies from difference sources. This is required for both load balancing and reducing energy costs. Adding more complexity, energy companies must also comply with relevant government and industry regulations.

It’s not that the industry hasn’t tried COTS software. It has, but with generally poor results. These software packages have fallen short of expectations, often costing more than they should while failing to deliver their designated outcomes. And regional differences that should be baked in, such as state and national regulations, are all too often left out.

In addition, many COTS planning tools give only mathematical outputs. Business users are expected to use these output values, interpret them and then use the results for decision making. That may sound good on paper, but it rarely works out well in the real world.

What’s needed instead is the ability to integrate these mathematical outputs and assimilate them into business processes — both upstream and downstream — while also offering possible scenarios. This, in turn, requires a new type of analytics solution, one that can “operationalize” an entire ecosystem and offer outputs that are not just mathematical, but also prescriptive.

Data deluge

These new solutions are needed, in part, because alternative-energy companies now find themselves handling extraordinarily large amounts of data — an activity far outside their areas of expertise.

What’s more, this data volume will only increase as the industry itself grows. Windpower generation is expected to double globally over the next five years, according to the Global Wind Energy Council, an international trade association. Not coincidentally, prices for wind and solar energy are expected to drop by nearly 60 percent by 2025. Wind power alone is expected to meet nearly 20 percent of the world’s energy needs by 2030.

All of this activity will mean more data. The data will come from wind farms, power grids, generators, transmission centers, distribution points, government regulators and others. Energy providers, in turn, will be required to ingest all this data, integrate its complex data points, and ensure seamless energy supply and forecasts. On top of all that, these providers will also need to share data with suppliers of more traditional energy sources, and then integrate it for effective load balancing.

Forecasting energy

For an example of how analytics can help the alternative-energy industry, consider a recent project completed by DXC Technology for one of our customers, a major distributor of renewable energy. This company came to us with three main objectives:

• Forecasting: The company needed the ability to accurately forecast how much energy its plants could generate on both a short-term (75 minutes) and long-term (48 hours) basis.
• Load balancing: It also needed to work closely with providers of other forms of energy to ensure an adequate energy supply.
• Cost control: Management called for new ways to lower the company’s costs for producing energy.

In addition, we supplemented this list with two other, related goals: Improve the customer’s forecast accuracy by 3 to 5 percent. And, reduce its overall forecast-error magnitude by half.

However, once we began work on the project, we discovered four serious issues with the company’s data:

• A lack of relevant features: Traditionally, energy suppliers rely on several weather conditions to forecast wind power, including wind speed, wind direction, air pressure and relative humidity. However, the renewable energy distributor was able to supply only one: wind speed, both speed that was forecast and speed that was recorded. Therefore, the model would have to be able to operate with only this one input. Also, the lack of other weather inputs could not be allowed to impair the forecast’s accuracy.
• Increased complexity: Our original plan called for the company to provide wind speed measurements at the farm level. The idea was that an overall forecast could then be based on an aggregation of the components. However, the company was unable to provide wind speed data at the farm level. Instead, it could provide only summarized data at the transformer level — that is, where each transformer station has one or more turbines feeding into it. As a result, the model would be unable to determine which turbines were causing the maximum variance. This presented a major challenge.
• Inaccurate wind speed measurements: Ideally the weather stations — including forecasting and recording weather stations, as well as wind farms — would have been co-located for the most accurate predictions possible. However, in many instances the weather stations and wind farms were as far as 35 miles apart. With that great distance came equally great issues with both measurement reliability and accuracy. This increased the project’s complexity, requiring us to bring additional aspects into the model.
• Inconsistent patterns: The laws of physics dictate that wind speed and windpower outputs are directly proportional to each other. That is, the more wind, the more power generated. However, we found many anomalies in the customer’s setup. At several locations, even with adequate wind speed, the actual power output was nil. Upon investigation, we discovered that the main culprits were both scheduled and unscheduled maintenance work. As a result, our model would need to be “trained” to deal with situations where maintenance was either scheduled or actually being done.

Dual approach

To create forecasting models for the company, we opted to divide the overall project into two parts: short-term and long-term forecasts. Once each part was completed, we could integrate them for a total solution.

Since we knew that the short-term forecasting engine would rely on long-term forecasts, we had to build the long-term system first. We began with four long-term datasets: forecast wind speed, actual wind speed, 18-month historical wind-power data and geospatial data.

We initially started building the forecasting models using time series regression. However, due to the underlying data challenges, the output was unsatisfactory. To overcome these challenges, we developed a long-term forecasting engine with two main components: Scaler and Optimizer. As shown in Figure 1, the Scaler was introduced to handle inaccurate wind speeds, and the Optimizer was added to handle inconsistent wind-power patterns.

However, even with these two additional components, the model needed further tuning. To address this and reduce deviations, we used several scaling and optimization techniques. These helped us reduce the error deviation by as much as 50 percent.

For short-term forecasts, we began with the same four datasets we used for the longterm system: forecast wind speed, actual wind speed, 18-month historical wind-power data, and geospatial data.

Next, we designed an innovative methodology to generate the short-term forecast. We leveraged the data of both recorded and forecast wind power to forecast for the next 75 minutes. For this, we built a background scheduler (also known as a “daemon”) and a short-term forecasting engine to run every 15 minutes to optimize the forecast for the next 75 minutes, as shown in Figure 2.

We then designed the short-term forecasting engine to leverage the in-house long-term forecast wind power and the recorded wind power in order to forecast for the next 75 minutes. The engine’s efficiency was ensured by optimizing the forecast values of the next 75 minutes, based on the deviation between the forecast for the previous 60 minutes and the actual recorded wind power. So, the run time is in microseconds, and the forecasting is in real time.

Impressive improvements

Once our solution was packaged and ready, the energy distribution company was able to run the model. Together, we evaluated both the long-term and short-term models against the benchmark data using historical data across varying seasons, days and transformers. Our solution for forecasting wind power using only wind speed was able to beat the benchmark 95 percent of the time. It also improved the forecast accuracy by 5 percent. And deviation from the actual results, compared with the benchmark, was reduced by almost 50 percent.

Also, our forecasting engine was able to predict the next 48 hours of wind-power energy with wind speed as the only feature. When traditional methods did not suffice, our Scaler and Optimizer enabled higher-accuracy predictions.

Could this analytics approach to forecasting be applied to other types of alternativeenergy suppliers? Yes. Whether offering wind, solar or hydropower, these companies require tailor-made analytics to handle their huge and fast-growing quantities of data. Forecasting in an environment too complex for off-the-shelf solutions is a job for tailored analytics.

Read the full paper and contact us to learn more.