Role of AI in Solar Asset Management

Artificial Intelligence (AI) has emerged as a powerful tool for analysis of Solar Plant operational data.   In the past, operational managers were hamstrung due to sketchy knowledge of statistics.  Many could not go beyond mean, median, standard deviation and variance which has limited utility in a sector where one has to deal with probability.  While writing this piece I am also aware that it is the month of the year in which Operational Managers are required to give a monthly, quarterly and annual generation and revenue budget of each plant.  In a portfolio having 50-60 plants of different sizes, it becomes a mammoth exercise, and they spend considerable time without achieving sufficient accuracy.  While monthly reviews are done internally with immediate superior, quarterly and annual reviews take place by an appointed body by the business.   Many have faced such situations where the actual numbers are far from the estimated numbers and invariably actual ones are lower than budgeted.  Mangers scamper for some reason to explain this gap and often attribute it to lower than budgeted irradiation or dust in the atmosphere or higher soiling loss.  Some prophets of doom also suspect the excessive degradation of modules, and incorrect measurement of irradiation by pyranometer.  This saga goes on for month to month, quarter to quarter and year to year.  Often people do not have time to analyze data objectively and even if they have time, they are not equipped with the right tools to deal with it. 

Since there is a general lack of awareness and analysis is devoid of visible action, it is often looked down upon and person doing it is termed a knowledge terrorist whose job is to waylay people from main track of action.  However, to solve any problem, say lower than estimated generation, one needs to estimate the weight of each parameter contributing to the gap and accordingly decide action based on 80/20 rule.  80/20 rule is fine but suppose lower irradiation happens to be one among the major factors responsible for lower generation, which is beyond your control, there are few options.  However, one can derive some learnings for future plants.  In many collocated plants, there is a micro-climate above the plant which is different than the surrounding area.  In such cases, a correlation between meteorological data with the air quality index, wind velocity and wind direction, etc. can be used to moderate the GHI.  Similarly, if 2-3 years of performance data show variance with the estimated number, there is no point in taking the design value as the estimated irradiation and generation number and continue to play broken record.  A more prudent approach is to set a new monthly, quarterly and annual target using historical data and track performance with respect to same. I have also experienced that generally variable salary of the operational team is linked with the revenue numbers and it is an absolute single number and anything above this number puts them in outperform category and below it in the underperformance category.  Since irradiation is probabilistic in nature and varies in a band of ±3-4%, how rational it is to have one single number for ‘meets performance’ criterion.  It is required to be a band and anything below the band is under performance and above it over performance.  I confess that even in my professional career I have committed similar mistakes but now with the availability of more powerful tools the philosophy of target setting can be modified.

I did an analysis for a plant located in central India and it took only 15 mins to arrive at a reasonable target.  The output runs in several pages but for brevity, I am just summarizing the old and AI suggested target which was validated for its accuracy for the next year’s generation data.  The validation showed +4.32% error in annual irradiation and -0.46% in annual generation.  Monthly absolute irradiation and generation errors were 6.67% and 4.65%, respectively.  It also gives month-wise errors and identifies months in which error is maximum and minimum.