Interview with Qiuhua Huang & Jason Hou- Application of Machine Learning in Power Systems – Part 2
Qiuhua Huang received his B.Eng. and M.S. degree in electrical engineering from South China University of Technology, Guangzhou, China, in 2009 and 2012, respectively, and his Ph.D. degree in electrical engineering from Arizona State University, Tempe, AZ, USA, in 2016. Qiuhua Huang is currently a power system research engineer in the Electricity Infrastructure group, Pacific Northwest National Laboratory, Richland, WA, USA. His research interests include power system modeling, simulation and control, transactive energy, and application of advanced computing and machine learning technologies in power systems. Currently, he is the principal investigator/project manager of several DOE funded projects. He is co-chair of the “Deep Learning and Smart Grid Applications” panel session at PES GM 2018. He is an Associated Editor of CSEE Journal of Power and Energy Systems.
Dr. Z. Jason Hou received his Ph.D. from U.C. Berkeley, and is currently a senior data scientist and statistician at Pacific Northwest National Laboratory. He has been well recognized for leading pioneering work in developing and applying advanced machine learning, uncertainty quantification, and extreme events analysis, in the areas of earth, energy, and environmental systems. Dr. Hou's research broadly cuts across areas in stochastic operation and planning of energy systems, extreme events in earth systems, carbon sequestration, oil/gas exploration, and environmental remediation.
In this interview, Qiuhua & Jason answer questions from Part 2 of their webinar, Application of Machine Learning in Power Systems", originally presented on Nov 1st, 2018. For more details regarding these questions, please view his webinar on-demand on the IEEE SG Resource Center.
Is it possible to use machine learning (ML) for power system state estimation?
What is the difference between path flows and line flows?
Line flow is the power flow of single transmission line. A transmission path (or channel) usually comprises multiple transmission lines connecting two areas, responsible of moving bulk power from one area to another area. Thus, the path flow is the total flows of all the lines in the path.
Is there any plan using AI for network model management?
First, we are not fully sure about what “network model management” really means here. If it is related to data or case management for network models, AI could play a role, for example, it be used for detecting outliners or errors in the model data or study cases.
Are the data available for public mining and analyzing?
Several datasets were discussed in the presentation. Publically available data includes those from US Energy Information Administration (EIA) and National Climatic Data Center (NCDC). Some other data (e.g., PMU and system oscillation) are of limited distribution.
Is any interlink between how we use the big data for machine learning/AI?
Machine learning can be considered a one step further than big data analytics as it can learn from the existing data in addition to extracting patterns. ML can be applied to small datasets as well. Big data usually have great spatial, temporal, and/or scenario coverage, which help reduce the overfitting issues when machine learning techniques are applied.
Can you share insight of data preparation for performing training for deep Neural Network?
For deep neural network learning, significant amount of data is usually needed to cover as many scenarios as possible, more than what is needed by other simpler machine learning approaches, due to its overfitting issues and lack of predictability of unseen patterns/events. Depending on the hypotheses that one wants to test and questions to answer, the data should be representative of most possible scenarios - note that the data preparation efforts can be reduced via smart sampling of experimental design and/or integration of reinforcement learning.
How is the performance of the Deep Reinforcement Learning in a partial observable scenario? (i.e. Not enough meters/sensors to cover the grid).
Deep reinforcement learning has been applied to many partial observable scenarios and achieved good results. The grid emergency control case discussed in part 2 showed good performance in a partial observable scenario, where only a portion of the system of interest are monitored and only bus voltage and loads are monitored and used as input to the deep neural network. A recent successful application story is AlphaStar, developed by DeepMind, beat human professional players in the StarCraft 2 game, which is partially observable.
Is it a valid use case to use machine learning to predict the outages?
Yes extreme events in different systems can be predicted using machine learning, and the key to find a reasonable set of factors (e.g., weather extremes, socioeconomic events, physical disruptions). The more cases are being included in the model training, the more confidence we have about the prediction accuracy. Many classification and regression techniques are applicable for the purpose. A recent study based on historical outage data from Bonneville Power Administration (BPA) paper is reported in this paper.
What is the performance of machine learning in solving OPF in comparison with mathematical algorithms?
The mathematical algorithm – based numerical models are already an approximation of the system, while the machine learning models are a further approximation of the numerical models, by improving the computational efficiency while sacrificing certain accuracy due to the nonlinearity and dynamic behaviors of the system. A simple answer is that the machine learning models are more efficient but less accurate, but its accuracy can be improved by choosing the most appropriate algorithms via model selection.
How does AI incorporate unseen system operating conditions?
It is desired that one has all necessary data such that AI can make predictions only about known situations without extrapolation. But because of time and resource limitations, it is practically impossible to gather all these data. If unknown situations occurs frequently, iterative or reinforcement learning can be used to make the learning process adaptive to new data and conditions, with the learned model updated as new data comes in. The new cases might be classified into existing groups or assigned a new label or labels.
Which Deep learning platform/software was used for “accurate day ahead forecasting" in point 4?
The software has three core algorithms: a cloudera Hadoop 5.9 HIVE database system with automatic data harvesting from NOAA and EIA websites; built-in data cleaning (e.g., time stamp correction, outlier removal) and exploratory data analysis subroutines written in R; and grid stress prediction with artificial neural network (ANN) written in MATLAB. All the programs are launched daily for day-ahead grid stress forecast with an automatic launcher.
How can ML can be applied in adaptive power network protection?
We don’t have much research experience in this area yet. But in general, machine learning could help improve the adaptiveness of power system protection from the following perspectives: 1) scenario-based protection setting instead of one fixed setting for all scenarios; 2) exploiting some abstract or high-level features extracted from input signals with machine learning techniques, such as long short-term memory neural network (LSTM), to identify and distinguish different faults and/or operation conditions.
Can you have a confidence interval (risk measurement) for the prediction of the neural networks?
If uncertainty quantification is key to a study, multiple ensemble machine learning approaches including Gaussian mixture models are recommended. For neural network analysis, one idea is to use Monte Carlo drop-out technique to provide multiple predictions for quantifying uncertainty with multiple networks.
Is there any negative effects or disadvantages of using machine learning in power systems?
All methods have pros and cons, with no exception to machine learning approaches. Today, main disadvantages of using machine learning in power systems are 1) there are inadequate quality and length of data obtained from real-world power systems (many existing work or publications are based on simulation or synthetic dataset); 2) many machine learning methods may not produce satisfactory results if they are used out of the box with default parameters (i.e., performance is sensitive to the parameters or hyperparameters). Looking forward, both could be addressed with more training data and better understanding of the most appropriate algorithms for different the power system applications.
How do you implement the angular prediction algorithm using machine learning?
We are not sure about which part of the presentation this question was referring to.
Is there any reference for machine learning application in adaptive power network protection?
We didn’t have direct research experience in this area. Some interesting publications can be considered (e.g., Dubey et al 2016 in IET Gener. Trans. Distrib; Tang and Yang 2017 in Energies; Tomin et al 2016 in IFAC).
What are some potential applications of ML in Microgrid control?
As presented in the part 2, one potential area is (deep) reinforcement learning for power system control, including Microgrid control. Also see Chaouachi et al 2013 in IEEE Transactions on Industrial Electronics.
Can you recommend some open source applications for the smart grid development?
Here is a list of open-source applications:
- Reinforcement learning for grid control: https://github.com/RLGC-Project/RLGC
- TensorFlow Based Machine Learning For Power System Analysis: https://github.com/interpss/DeepMachineLearning
- Local Feature Sufficiency Exploration for Predicting Security-constrained Generation Dispatch in Multi-Area Power Systems: https://github.com/sunyx1223/GridOptimization
- Model-Free Renewables Scenario Generation Using Generative Adversarial Networks: https://github.com/chennnnnyize/Renewables_Scenario_Gen_GAN
- Short-term load forecasting with deep residual networks: https://github.com/yalickj/load-forecasting-resnet
Any model to show how ISOs should keep and store time series data from PMUs for AI?
Please refer to the IEEE Standard for Synchrophasor Data Transfer IEEE C37.118.2-2011, and NASPI reports https://www.naspi.org/reference-documents.
How can we relate artificial intelligence and machine learning?
AI is much broader than machine learning, and machine learning is generally regarded as subset of AI. For more details, one can refer to this article: https://www.datasciencecentral.com/profiles/blogs/artificial-intelligence-vs-machine-learning-vs-deep-learning
Is Recurrent Neural Networks method efficient for Power Quality Classification?
We don’t have direct research experience to answer it. But in practice, multiple classification approaches should be compared and integrated as the most appropriate approaches may vary from case to case.
When talking about real-time application of machine learning, how fast is it? What sampling time we are talking about?
There are at least two types of “real-time” applications. One type is based on a machine learning model which is trained off-line, and it could be run as fast as needed, and the sampling time will depend on the application. For PMU anomaly detection, the PMU sampling rate is used. Another type is that learning happens in a real-time environment. The speed would be somehow constrained by the computing power and IO. Different learning algorithms or design parameters can be adopted depending on whether we are targeting a predictive model within minutes, seconds, or sub-seconds.
Do you think that AI can be applied in real protection relay problem?
This is one of the active and promising research areas in the power domain. There have been a lot of research efforts on applying CNN or LSTM for power system protection/relay.
Is that true that most of the current ML algorithms proposed in power systems are vulnerable to adversarial examples, which are maliciously crafted input data?
This is a very good question, and it is also related to one important research topic-- adversarial machine learning. As far as I know, this is an open question and much more research work should be done to understand it. Here is one recent research paper in the power domain: https://arxiv.org/pdf/1808.08197.pdf
Do we need to scale or make all data in per unit?
Normalization is not a must, but in general a good practice particularly when neural-network-based technique is used.
Can you compare the efficiency of your approach for interchange flow prediction comparing to previously developed by PNNL tool (DINA) based on statistical analysis?
This is a good question and suggestion. While both approaches are targeting at interchange flow prediction. The Bayesian Model Averaging (BMA) method presented in part 2 is basically a deterministic method, while DINA provides probabilistic estimation of the secure range of the net interchange of a BA. Thus, it is difficult to compare them directly.
In one of the Neural Networks you had 2 hidden layers. How do you choose the number of hidden layers?
The number of hidden layers is important aspect of the neural network architecture. Unfortunately, there is no easy or straightforward way to determine this. In general, based on the complexity of the problem, one should start with a reasonably small number and increase it as necessary based on the training and testing results, aiming at improved accuracy while minimizing overfitting.
Does Neural Network training results provide better results if data scaled to 0 to 1 instead of actual MegaWatt or Ampere?
Data scaling may or may not improve the results, but is expected to facilitate the learning process by avoiding potential numerical issues in the neural network training. One may see quick saturation of the hidden units and near zero gradients without data scaling.
To view past interviews, please visit the IEEE Smart Grid Resource Center.