Modern power systems are experiencing fundamental changes that are driven by global warming policies, market forces, and the advancement of technology.
They are, at the same time, facing multiple challenges on different fronts,more so in our continent. Assigned East and horn of africa,i embarked on a research journey to assess stability on our power systems across several countries in EA and HoA. somalia and Kenya were my field study grounds.
Challenges and Future Research Opportunities
The novel paradigm of employing machine and deep learning in power system TSA has proven popular among researchers and has shown great promise on benchmark test cases. However, building the next generation of TSA tools, using data mining and artificial intelligence, is still a work in progress. Namely, stepping from benchmarks into the real- world power systems is often plagued with difficulties emanating from system complexity and unforeseen circumstances. Considering the importance of power systems, there is still ample need for research that corroborates the quality and robustness of this data-driven AI approach to solving the TSA problem. Hence, Appendix A provides another brief overview of selected research concerning different state-of-the-art AI approaches to power system TSA. This overview presents a snapshot of the current state of affairs, hints at open challenges, and at the same time, points toward future research opportunities.
In the domain of dataset building,several challenges facing the research community are identified: (1) open sourcing datasets from power system WAMS measurements,
(2) addressing potential security concerns associated with WAMS data,
(3) open sourc- ing existing benchmark test cases,
(4) providing a unified and consistent set of benchmark test cases featuring power systems of varying sizes and levels of RES penetration,
(5) providing benchmark test cases featuring hybrid AC/DC power grids,
(6) providing standardized simulated environments of power systems for the reinforcement learning.
These points deal with standardization of benchmark test cases and bringing them closer to the expected future levels of RES penetration, as well as introducing the hybrid AC/DC power grids. It also deals with bringing standard simulated environments for RL (i.e., something such as OpenAI Gym (https://gym.openai.com, accessed on 18 November 2021) for power systems). The hard work of building some of these components is already under way, with the support of the LF Energy (https://www.lfenergy.org, accessed on 18 November 2021) umbrella organization of open-source power system projects. In the domain of data processing pipelines, challenges include:
(1) automatic data labeling,
(2) cre- ative features engineering for real-time performance,
(3) dealing with the class imbalance problem,
(4) dealing with missing data,
(5) dealing with data drift,
(6) using embedding as a features space reduction, and (7) using autoencoders with unsupervised learning for dimensionality reduction.
In the domain of model building, several issues have been identified with some of the DL-based image classifiers when applied to the power system TSA. In addition, training of deep learning models, in general, is associated with its own challenges: proper layer initialization, learning rate scheduling, convergence, overfitting, vanishing gradients, forgetfulness, dead neurons, long training times, and others.
Reinforcement learning models, at the same time, are even far more difficult to train. Furthermore, RL models tend to be “brittle” and may exhibit unexpected behavior.
Tackling these different challenges, at the same time, presents new research opportunities. In the domain of synthetic data generation from benchmark test cases, future research may address the following issues for stress-testing the existing models:
(1) introducing different types of noise and measurement errors into datasets,
(2) introducing different level of class imbalances into datasets,
(3) introducing different levels of RES penetration into the IEEE New England 39-bus test case,
(4) using deep learning for features extrac- tion from time-domain signals,
(5) artificial features engineering,
(6) speeding up numerical simulations of benchmark test cases with parallel processing or by other means,
(7) introducing hybrid AC/DC power grid test cases for transient stability analysis.
In the domain of model building, future research opportunities arise from applying different deep learning models to the power system TSA problem. Employing stacked autoencoders, transfer learning, attention mechanisms, graph neural networks, and other state-of-the-art deep learning architectures is seen as a way forward. Particularly important would be applications of deep learning architectures designed specifically for processing long time series data, image classification, and analyzing graph structures. This is generally seen as a major area of deep learning research, where any novel architecture in this area may be tested on the TSA problem as well. It would also be interesting to see if three-phase signals can be exploited as RGB channels in convolutional layers, with cross-learning between channels/phases. There are research opportunities in devising novel and better ways of converting multivariate TSA signals into images for use with very advanced deep learning image classifiers.
Another major issue with deep learning is a general lack of model interpretability, cou- pled with the difficulty of understanding the model’s decision-making process. However, there is active research in this area as well.
All major deep learning frameworks allow fast model prototyping and training on powerful distributed hardware architectures in the cloud. This levels the playing field for researchers and lowers the barrier for entry. When it comes to model maintenance, TSA is associated with a data drift phenomenon emanating from steadily increasing RES penetration. Efficiently dealing with data drift is a point of concern for model serving, along with issues of model latency and throughput. Namely, model prediction serving performance (i.e., latency) needs to be in real time for maintaining transient stability of power systems following a disturbance. This limits the computational time available for pipeline processing and may impose certain constraints on its design.
5. Conclusions
The deep learning domain is, generally speaking, still a very active area of research, with several promising avenues: different RNN network architectures (with LSTM or GRU layers), stacked autoencoders, transformers, attention mechanism, transfer learning, and GAN networks, to name a few prominent ones. Deep learning models specifically designed to learn from multivariate long time series data are particularly interesting for applications to the power system TSA problem. Exploiting the graph-like structure of the power systems with the use of graph neural networks is another promising research direction. Furthermore, deep learning models that use unlabeled data (such as stacked autoencoders) are very attractive, since data labeling can be time-consuming and necessitates applying humans with domain expert knowledge.
Reinforcement learning, as a special subset of deep learning, is yet another approach that shows early promising signs. However, the training of deep reinforcement learning models is notoriously difficult. The landscape of reinforcement learning is rapidly expand- ing, which offers ample research opportunities for integrating it with transient stability assessment, particularly in the area of power system control following a disturbance. This is still a relatively nascent, but very promising area of research.
The importance of the electrical power system to society mandates that further con- vincing results be provided in order to corroborate the stability and robustness of all these various AI approaches. Furthermore, additional and extensive models stress-testing, with different levels of data corruption, is warranted. Potential security issues connected with the wide dissemination of actual WAMS/PMU data need to be addressed as well. This creates space for new research outputs that can fill this gap and increase the overall confi- dence of the entire community in this new technology, for its safe future deployment across power systems.
Author : Samwel Kariuki
Date: 29th Feb 2024