Solar Farm Power Output Prediction Model Using Deep Learning

Solar farm power output prediction model using deep learning

Objective

To develop a solution that involves processing data from varied sources and create a prediction model to forecast the power output of the solar farm.

Business problem

To help the operations of solar farm know about the output that it would produce in next 4-6 hours. This is short duration forecast and usually determined for a smaller region of farm ecosystem. The inputs were derived from the images of sky-camera installed in the farm and weather data coming from the trusted real time sources. The first part of the challenge was to test the model to predict the solar irradiance given the images and weather conditions. Once the irradiance model was completed and tested with independent data, the farm power output prediction was implemented.

Solution Overview

Creation of data pipelines along with setting the training compute targets on NVIDIA GPU.
The preprocessing of image was completed with appropriate filters and then a Convolutional Neural Network along with a Sequence Model was implemented.
This technic of image sequencing is advantageous for Video Tagging and Processing and same has been leveraged extract the information from Sky Images and perform feature extraction for the final model.
The first version of the model has been deployed as a pilot and has reported an accuracy of 82% which is improved output in terms of forecasted values.
The existing models in industry that either work on LSTM alone or Auto-Regressive models have reported accuracy of about 70% and also not considering the solar irradiance.
The technology success of the project is a model that is scalable to other regions and only feature dependency of solar irradiance that can be retrained.
This model is being refined as AKIRA Solution for the Solar PV Market and aims to partner with more organizations active in the same solution space.

Algorithm Block Structure

Business Outcome

Shell’s Solar Vertical has reported a 30% improvement in Power Procurement from the grid as our solution has empowered them to decide in advance for power buy or sell decision and largely create an ecosystem of multiple solar farms with near to grid parity economy.

Solution Architecture

Solution Brief

The implementation is based upon CNN-LSTM processing of the images and other tabular data to implement the classification of solar irradiance.
Model predicts the irradiance and estimates the power output of the solar farm as solar irradiance and farm power output is directly proportional.
Enhancing the model to predict the Power output based upon the Image Data, Weather Conditions and Previous output without calculating irradiance part.
Reliable forecast is the first step to optimal integration of renewable power generation into traditional power networks. Therefore, in a bid to repugnant the adverse impacts of renewable power integration with the grid, we have focused on developing model that provides prediction with an accuracy that could enable the operation of such plants like conventional ones. Also, intra-hour forecasts of fast ramp rates in solar irradiance is pertinent for intra-day electricity market participation.

Goals

This solution is being tested in the field and our team is continuously improving the model in bid to get to the best generalized deep learning solution that can be deployed easily and economic to maintain so that upcoming energy start-ups can partner with us.
The sustainable and clean energy alternative is dependent upon the wholesale PV Solar market and our solution is improving the way our Solar Markets manage their production.