Akira Insights - Developing a Cutting-Edge ML Model to Forecast Solar Power Production

Names:

Shivam Rai, Practice Head – Data Science
Anant Shukla, Senior Analyst
Aditya Aranya, Analyst

Country: India

Organization: Akira Insights Pvt Ltd, India

Akira Insights serves the rising end-to-end data & analytics needs of organizations in domestic and global markets. We help enterprises plan and execute their transformation to a truly data-driven business by enabling them to understand past events, predict future trends, and prescribe the optimal course of actions

Awards Categories:

• Data Science for Good
• Most Impactful Transformation Story
• Excellence in Research
• Partner Acceleration

Business Challenge:

The solution aims to address the solar power forecast for wholesale photovoltaic market segment with average 100 MW capacity. The renewable energy sector has witnessed a multi-fold increase in generation capacity and expected wholesale solar photovoltaic market size by 2025 will be USD 150 Billion (according to IAE, Fortune Business Insights).

Energy producers are dependent on the accuracy of the forecast to meet the grid supply and demand, as well as to make decisions on the operational front. They need to calculate the forecast and make a decision on buying from real-time sources and selling to regional grids in case of lesser power generation. There is a penalty component on not meeting the average power committed to the grid.

To reduce cost and improve efficiency we developed a deep learning model that consumes sky-images, weather data, and sensors data to produce the solar irradiance forecast. The power generated from each site is directly proportional to the irradiance value.

The solution makes a prediction every hour to estimate the value of power being generated. The model is capable of making such forecast for an extended interval of four hours. Having an hourly forecast enables us to act in case of any sudden dip in irradiance. The model also allows us to cater the solar farm sites that encounter patchy clouds and usually in the regions of China, Australia, and Southern Europe.

There are multiple methods used to forecast the irradiance but most of them are time series models with a lag of 24 to 96 hours. With a sky-image and weather data as input, the model can predict the next 1-4 hours of output.

Designing and implementing such a cutting-edge idea was complex and we had to go a lot of back-and-forth with data selection, pre-processing, and algorithm selection. This requires both speedy action and experimentation. Another challenge while working with a niche model is to bring domain-technology experts together with the data scientist group and collaborate.

To develop such a model, we required domain expertise, hence partnered with an energy giant to test our model at various sites and a baseline model is deployed for testing. The improvisation on algorithm and operational aspects is in progress.

Business Solution:

Working with Dataiku has been beneficial in developing a working model given the complexity of the data at hand. We decided to leverage Dataiku to implement the solution, as we needed a platform that could support rapid experimentation and prototyping.

The level of collaboration a team can achieve with Dataiku is another advantage. It’s a lot easier to reproduce the experiment results, compare different models, and proceed to further analysis. The support to integrate with components like MLflow and Container services for model hosting is also productivity enhancer. The whole workflow, from design to model hosting model and entering into reiteration, is made super easy.

Business Area Enhanced: Product & Service Development

Use Case Stage: In Production

Value Generated:

Model recorded a Mean Absolute Percentage Error (MAPE) 12-14% on solar irradiance prediction. According to studies on forecasting economy of solar power, any improvement above the persistence forecast saves 50% of the penalty cost, and 1% improvement on forecast accuracy beyond a level of 80% empirically generates an additional USD 30k-50k per year - considering a 100 MW solar plant with average zonal tariff in US markets.

A 100 MW Solar plant saves 9,000 tons CO2 emissions, compared to conventional methods of burning gas or coal – hence our model is a step to contribute to success of mid-size and wholesale market segment in solar power.

Value Brought by Dataiku:

Dataiku is an end-to end ML platform and the level of collaboration helps teams achieve results in a cross-functional environment. The explainability and experiment reproducibility are crucial features for the success of our project, alongside the range of algorithms and platform integrations that Dataiku offers. We created a baseline structure for data pipelines, preprocessing and model training, and we could also launch our model for inference and be embedded with other testing apps.

The upskilling effort required is lesser and learning is easy for a non-ML practitioner. For team Akira it was an ML made easy experience – fulfilling the Everyday AI mission.

Value Type:

• Reduce cost
• Reduce risk
• Save time