To risk or not to risk? that is the question – AI is the answer

This is where AI and Machine Learning become necessary for utilities.  AI/ML enables: predictive maintenance, streamlined operations, anticipation of equipment failure, more precise disruptive weather event forecasting, etc. [1]   It provides for pro-active management, including grow-out of an energy generating portfolio.  For one simple example of cost-cutting vs. pro-active management, consider vegetation control.  There are now almost 2 dozen drone service companies, which can provide drone equipment or inspection, so that monitoring of vegetation can become real-time, as-needed, instead of by a fixed schedule which may be too often (more expensive) or too infrequent (expensive if a fire results).  The drone data is however big-data and requires AI/ML types of algorithms to extract the image-based information of where issues are developing, and trimming needed.  Importantly, implementing this pro-active mode of AI/ML & drone data is a low risk compared with the reactive response of power shutdowns or to fires which can be devastating.  So, the risk of NOT using AI/ML and drones becomes much higher.   As another example, transformers can be monitored for operating signatures, which again with the right AI/ML analysis can indicate incipient issues that can be caught before a breakdown or an outage.  These are just some of the modes where the risk of implementing AI/ML and developing new approaches are less than the other risks that can occur by using standard operating procedures.  The AI/ML can mitigate risks leading to an overall reduction and many examples are already out there. [2]

TRUST

One of the key aspects to relying on AI/ML for any operational or forecasting aspect comes down to trust.  Can you trust the results?  Any complex problem can be addressed with lots of AI/ML resources and model training, but ultimately can you accept the results as accurate and bank on them.  The answer is “it depends”.  There are a lot of issues with data sets, training the algorithms, and the selection of algorithms.  Two key issues include (1) over-fitting the data, or (2) tuning to the noise.  These are real concerns and need to be addressed by an adequately sophisticated AI/ML approach.  However, it is indeed possible to get results with high accuracy which can be “trusted”.  As an example of some of our prior work, figure 1 shows the accuracy of prediction of the electricity demand for a utility over an extended period.[2]

In this engagement with Taiwan Power Company, the Ghiassi AI/ML modeling gave highly accurate Predictive Analytics with mean absolute percent error below 1%,  which is better than the standard approaches of multiple linear regression, and of traditional Advanced Neural Nets, by >1000% & >300% respectively.  In this example the accuracy is >99% in terms of the forecasted vs the actual data, and with the Ghiassi algorithm analysis there is both no over-fitting, and no tuning to the noise.  (See the last section for a discussion on achieving 100%.)

Figure 1.  Excellent Forecast of Total Power Sales for an electrical utility, by Ghiassi AI/ML [3]

Figure 2.  Another example of >99% forecast accuracy with the Ghiassi AI/ML [4]

TECHNICAL AI/ML DISCUSSION

A bit on the technical side, the Ghiassi (author’s) algorithms performed better than alternatives without over-fitting while data sets for training were kept separate from the testing (out-of-sample) data set.  Current published research presents a systematic and critical review of forecasting methods used in 483 utility forecasting models and establishes that ANN is the most utilized approach and outperforms alternative statistical methods in forecasting key parameters such as e.g. energy demand.  The other approaches suffer from long training time, unstable optimization processes, or sensitivity to hyper-parameters.   Furthermore, in the ANN space, the Ghiassi AI/ML set has been shown to be superior to traditional ANNs (Velásquez 2012, Olssonz 2017).  According to recent studies: using accuracy metrics such as mean absolute percentage errors results, the Ghiassi model has better performance than the regression model, existing traditional ANN models, SVM with genetic algorithms (SVMG) model and SVM with immune algorithms (SVMIA) model.   (Wang 2010, Zheng 2014)   Current studies support the Ghiassi algorithm’s advantages. (e.g. Biswajit 2018)    

DEATH & TAXES 

Nothing is 100%, well except for death & taxes, although some even dispute both of those (but I digress).  And as our data is an example, AI/ML accuracy can be 99% and doesn’t need an IBM Watson level of capability.  This is the indication that adequate accuracy can and has been achieved without over-fitting.  With this kind of accuracy, improved implementation will grow with the learning curve, as the utilities develop a trust in the results and see the advantages for pro-active postures.  This is similar to many other industry learning curves and adoption.  The main takeaway is that it is now a bigger risk to not take the risk on AI/ML.

So not to be too morbid, Shakespeare’s Hamlet contemplates death and suicide, bemoaning the pain and unfairness of life but acknowledging that the alternative might be worse.  Doesn’t that sound a bit familiar? 

To risk, or not to risk.  That is indeed the question, but at least you know the answer (to within 99% certainty).

REFERENCES

[1] M. Ghiassi and A. Skumanich, "On the use of AI as a requirement for improved insolation forecasting accuracy to achieve optimized PV utilization" https://doi.org/10.1109/PVSC43889.2021.9518470

[2] A. Skumanich and M. Ghiassi, “Achieving significant cost reduction in solar through the targeted use of AI and Machine Learning”  https://doi.org/10.1109/PVSC43889.2021.9518872

[3] Ghiassi, M., et. al. (2006).  “Medium Term System Load Forecasting with a Dynamic Artificial Neural Network Model,” Electric Power Systems Research Vol. 76, 302-316

[4] Ghiassi, M., et. al. (2006).  “A Dynamic Artificial Neural Network Model for Forecasting Time Series Events”. International Journal of Forecasting, Vol. 21, 341.