PMU Tracker: A Visualization Platform for Epicentric
Event Propagation Analysis in the Power Grid

This work sits at the intersection of two primary areas: (1) visual analysis of the electric power grid, whereby domain operators and researchers visualize grid topology and data streams (including PMU data), and (2) time-varying topological visualization, where we consider networks with attribute values that change over time, and the topology of the network is an important consideration for analysis.

To help motivate our software development efforts, we surveyed three power system engineers that we have previously collaborated with to understand current challenges for visualizing and analyzing PMU-based power grid data. These industry professionals have extensive experience working for U.S. power companies (13, 15, and 32 years) in contingency analysis, systems protection, and control and coordination.

Table 1 lists several challenges explicitly mentioned by the practitioners for visualizing PMU data. While all three practitioners regularly use a combination of general-purpose scripting tools/languages and in-house/dedicated visualization software (RTDMS, ASPEN, PSS/E, etc.) to analyze both historical and real-time grid data, they each described event analysis workflows as “unintuitive.” Often, event analysis involved reviewing line charts of the PMU data and manually matching PMUs to their network/geographic locations. Event-related visualizations are generally manually created, such as via scripting languages (R and Python were commonly referenced). Such a manual workflow imposes a significant overhead on operators, limiting their ability to efficiently analyze anomalous data in the grid.

Based on the challenges listed in Table 1, we defined a set of task requirements that a visual analytics system can employ for analyzing and tracking the evolution of localized grid events.

Such a system could potentially benefit power systems engineers including those engaged in (i) contingency analysis, (ii) systems protection, control, and coordination, (iii) line engineering based on historical event data, (iv) handling transient and dynamic stability in the grid, and (v) the analysis of disturbance events, misoperation events, and synchrophasor data.

T1: Flexible and interactive data retrieval. Systems should provide a backend that supports real-time, customizable data retrieval to support flexible front-end exploration and enable early event flagging.

T2: Link events to PMUs. Localized events can be sourced to a specific region of the power grid network, either via automatic means [18, 13, 40] or by manually identifying one or more epicenter PMUs, which are adjacent to or at the event source, to begin tracking event propagation throughout the network.

T3: Support temporal analysis. As PMUs provide fine-grained data sampling (30 Hz or higher), subsecond events can be traced over the network, though sometimes events can have a duration effect of multiple hours. As a result, visualizations should flexibly support temporal analysis at multiple scales.

T4: Provide multiple perspectives for analyzing events. Providing multiple coordinated visualizations that support different data perspectives can provide a robust view for analysts who are identifying and analyzing an event. In particular, as many grid events begin as localized phenomena that spread throughout the network, including an epicentric view can potentially provide important analytic benefits to understand propagation. In our case, we incorporate views that support signal processing of temporal data streams, a global view of the grid network, a similarity-based plot of PMUs, and the novel dendrogram plot for epicentric analysis.

In contextualizing T1–T4 to existing grid software, to our knowledge no existing industry or research platforms wholly address all of these tasks. In this way, our system represents a scalable, end-to-end spatiotemporal solution for event localization and propagation tracing over a power grid.

PMU data is large-scale, noisy, and multi-attribute, but to support task T1, the retrieval of PMU data should be fast, flexible, and interactive. In this paper, we use a dataset that comprises $\sim$500 PMUs over a three-year period from a regional U.S. gas and electric company.¹¹1Access to this PMU dataset is provided under NDA, so we anonymize certain features in this paper, such as names, IDs, and locations, and only reference the approximate number of PMUs available. This dataset additionally contains a collection of operator reports about historical grid events, stored as unstructured text files.

We save the raw PMU data to disk on a compute cluster using the Parquet file format. Parquet is a columnar storage scheme based on an algorithm for record shredding and assembly [27]. It supports efficient and flexible compression and encoding. A Parquet file is stored as a series of blocks (row-groups), each of which in turn contains column groups; the values of the column groups are stored in contiguous memory locations. This transparent column-specific compression allows for high-performance data fetching compared to row-wise queries.

In our dataset, each Parquet file stores one attribute’s worth of data for all PMUs for one day (using 96 fifteen-minute groups), see Figure 1. The PMUs in our dataset record measurements at 30 Hz (during data dropout, null values are recorded), meaning each Parquet file is $\sim$500 columns $\times$ 2.592M rows. As the PMUs in our dataset record 18 attributes (with raw values stored as 64-bit floating-point values), each day contains 18 PMU files. In this manner, we optimize our storage for fast querying of PMUs for a single attribute of interest; for example, loading all $\sim$500 PMUs for a single day for one attribute requires just under six seconds (for one PMU: under two seconds). Query times scale approximately linearly as additional attributes are retrieved.

PMU data stored in Parquet files is compressed and decompressed automatically and on-the-fly using Gzip. The compressed size of our Parquet files can vary between attributes. For example, in compressed format, files for the Magnitude attribute are generally between 1.5–2.5 GB. When reading data from a Parquet file, uncompressing a row group results in reading approximately 20 MB of data. Storage-wise, each day’s set of 18 Parquet files is approximately 50–70 GB, with the full PMU dataset for three years being over 70 TB. There are current efforts to add an additional six years of PMU data to this dataset, which will equal nine years of data totaling over 200 TB in Parquet format.

Our PMU dataset additionally contains a collection of reports created by operators that describe grid events of interest, such as oscillations or faults. Reports are stored as unstructured text documents, meaning they do not provide any metadata explicitly linking the report to PMUs, substations, or locations in the grid (report timestamps are the only metadata). To link specific events to one or more epicenter PMUs, we use a regex matching schema between report text and PMU IDs, substation names (where the PMU is located), and their locations. If an association is found, the event is considered to originate from that PMU, or to originate closest to that linked PMU in the network.

One of the most promising applications of PMU technology is its usage in the detection and monitoring of power system oscillations. Generally, oscillations are triggered by system faults or generator misoperation, and they can be temporary (the oscillation is damped and dies off after a couple of seconds) or sustained. Sustained oscillations, also called forced oscillations, happen when a generator (or group of generators) has imperfect or inaccurate real-time control– it continuously injects active or reactive power that in turn excites the oscillatory modes. This type of oscillation can be stopped only when the generator is either taken offline or its controls are adjusted.

Oscillations can be observed from many different quantities such as voltage magnitude, frequency, and power injection (Figure 2). Oscillations propagate throughout the system and can be observed from many different buses around the source. An important task for power system operators is to localize the oscillation source– the bus at which the generator is exciting the system and causing the oscillatory behavior. Subsequently, corrective actions can be taken by the operators to fix the issue– i.e., controlling the generator or taking it offline.

We, therefore, incorporate epicenter-PMU identification and juxtaposition as a key functionality of PMU Tracker. We propose a method to localize the source of an oscillation, that fully integrates within our platform. When looking at bus voltage magnitudes during the oscillatory events in our dataset, it can be seen that the oscillations are the most prominent and clean at buses close to the source. As we get further, the voltages and the oscillations become noisier. This is due to the fact that as you get further from the source, the effect of the oscillation is damped and can be attributed to other phenomena and disturbances that are happening simultaneously in the system. Based on this observation, we perform an FFT analysis on the voltage and current signals of PMUs from different locations on the grid. By comparing the FFT magnitude at the oscillation frequency (Figure 3), we can get an idea of how close each bus is to the source. We illustrate this implementation within the platform in the following sections.

PMU Tracker’s interface for event analysis and oscillation localization is shown in Figure 4. The interface is designed to support tasks T1–T4 described in Section 2 and implements the FFT approach described in Section 4. It consists of seven linked panels (A)–(G) and supports linked highlighting: hovering on or selecting a PMU in one view highlights or selects the PMU in other views. A brief overview of the panels is as follows: (A) comprises the controls surrounding event selection and analysis parameters. (B) displays the grid network structure, as either a geographical or force-directed layout. (C) shows the main frequency of an oscillation occurring in the grid over a pre-specified time period. (D) and (E) visualize the FFT-characteristics of grid network PMUs that are either selected or unselected (respectively) for closer inspection. (F) uses a t-SNE to cluster PMUs based on a designated epicenter PMU for the historical/identified event being analyzed. (G) houses the novel epicentric dendrogram view, which allows operators to analyze the effects of an event as it propagates outwards from a source location.

At a high level, the interface can support the following workflow:

• •

  Users can begin either by selecting a historical grid event for analysis or browse the dataset by specifying parameters in (A). Available parameters include the time range to query the data, FFT calculation parameters (these help define the resolution for viewing the data), settings for flagging PMUs as potentially anomalous, and options for selecting a set of PMUs.

• •

  The grid network is displayed in (B)–(F), allowing users to analyze the state of the grid and compare the behaviors of flagged vs. unflagged PMUs

• •

  Epicentric event analysis is supported in the dendrogram view (G). PMUs can again be selected, and these are tracked to visualize how anomalies propagate through the grid network (from a source) during a system event.

To illustrate PMU Tracker’s available interactions, and demonstrate how it can be used to analyze anomalous behavior, we describe the interface’s features in conjunction with a usage scenario (shown in Figure 5) analyzing a historical grid oscillation event and discovering a previously-unidentified secondary event (demarcated as Scenario: … throughout this section’s text) from the perspective of a user, Martin, a transmission and distribution power engineer. As a note for the usage scenario, since our dataset contains historical (previously identified) events that we are analyzing, we start by selecting one in (A). However, in a real-time (or near-time) scenario (where no entry would be stored in the database), a user would begin by selecting the time frame of interest, along with a set of potentially-interesting PMUs to begin their analysis, in (A). A demo video of the interface is also available in the supplementary materials.

(A) Settings Panel. (a1) Users first select a grid event from a dropdown list, and its details are loaded. (a2) The timespan for FFT computation can be set along with the calculation window (2, 5, or 10 seconds), which slides across the specified start/end timestamps, and computes FFT values for the selected data attribute. (a3) To identify anomalous PMUs, the peak FFT magnitude value over all PMUs is extracted, and a threshold percentage is specified for this peak. PMUs with values above this threshold are flagged as anomalous. (a4) The flagged PMUs show (i) a selection icon, (ii) their ID, (iii) a color swatch showing their similarity to the epicenter PMU, and (iv) their FFT magnitude value at oscillation frequency. PMUs can be selected/de-selected from this list. (a5) Temporal analysis (T2) is supported via the time-scale slider. Scrubbing this widget dynamically updates the other panels to the current time value.

Scenario: Martin selects a System Voltage Oscillation event, with timestamp 20:44:00. The epicenter is located at PMU #122 at the Yearling substation, and the recorded duration is about 2 hours and is described to have an oscillation frequency of 2.4Hz. He decides to initially analyze the time period of 20:44:00-20:49:00 (about one hour after the start of the event), for VPm (the positive sequence voltage magnitude) with a 2 second FFT window. The threshold for anomalous PMU identification is set to 100% of the peak value, so only the epicenter PMU will be flagged. When finished, Martin clicks on the ‘submit’ button, and panels (B)-(F) are populated.

(B) Grid Network Panel. The power grid network topology is visualized using a node-link diagram. Edges denote power lines that connect buses, loads, plants, and/or transformers, which are visualized as rectangular nodes, and circular nodes denote PMUs. The network can be arranged using geographic locations, but to preserve anonymity Figure 4(A) shows a force-directed layout.

We note that the geographical grid network is highly interconnected, with multiple PMUs often existing on the same line. Real-world usage of tools like PMU Tracker primarily depends on this geographical view for event analysis, which follows industry standards and protocols. Event analysis involves working with a subset of PMUs of interest (that exhibit anomalous behavior) and not the juxtaposition of all network PMUs. The flagged PMUs (a4) provide context for this and are highlighted in the network view. Additionally, the white-to-purple color-map underlying the network visualizes a kernel density estimation of the peak FFT magnitude value for each PMU at the given timestep (T4). Such an approach follows industry standards, as a way to understand how an event’s magnitude spatially distributes over the network.

We provide standard interaction techniques here, including zooming, panning, node/edge filtering, and selecting PMUs via mouse click.

Scenario: On loading the information for the selected event (Figure 5(I) shows a zoomed-in view of the same), Martin sees in (B) that PMU #122, the only PMU that has been selected so far, is highlighted in the network, in red. The corresponding substation, Yearling, is colored green. All lines going out of Yearling are colored black, and the connected substations are marked in yellow. The rest of the network is rendered in gray and appears de-emphasized. Martin also notes that the underlying heatmap on the view has a relatively higher (purple) intensity compared to the rest of the power grid over the region surrounding the epicenter PMU (at the Yearling substation).

(C) Main Oscillation Frequency Panel. Oscillation frequency values are found by considering the position at which the maximum magnitude peak occurs at FFT calculated over every time step of the time-lapse period over all PMUs. This view allows the user to monitor the behavior of the oscillation over time: changes in the base frequency of the oscillation can indicate changes in the system response.

Scenario: Martin sees an initial oscillation frequency of $\sim$1 Hz shown in this view. He decides to examine a different time stamp from (A), as the event is described to have an oscillation frequency of 2.5 Hz. Martin scrubs the time slider control (a5) over the two-hour period following the event time stamp. He notes that throughout this duration in (B), the region surrounding the epicenter PMU (at the Yearling substation) has a relatively higher intensity compared to the rest of the power grid. Martin also finds that substations located 1-2 hops away from Yearling also show darker intensities compared to substations connected further away from the epicenter, and this peaks roughly an hour after the event time stamp. Martin selects 1-2 PMUs from each substation displaying these darker intensities for further review (a total of 20).

On viewing changes in (C), Martin finds that the main frequency of the oscillation gradually increases to peak at $\sim$2.5 Hz at around 21:45:00 (about one hour after the start of the event) over the VPm parameter. Following this, he sees the main frequency is maintained till 22:44:00, following which it declines to rest at $\sim$0.2 Hz over the next half hour. Based on these observations, Martin decides to, therefore, focus on the time period of 21:45:00-21:50:00 for the rest of the analysis. He resets the threshold (a4) to constitute 53% of the peak (0.0093) at this timestamp. He can observe that 9/20 selected PMUs have been flagged as crossing this threshold.

In the case of inter-area oscillations, system events along with poorly damped electric grid facilities can lead to groups of system units such as generators, transformers, etc. oscillating simultaneously at various regions in the grid. In this situation, main frequency monitoring is important, in order to gauge the impact of an event on the system.

(D) Selected PMU FFTs Panel. Selected PMUs are plotted to show their FFT magnitudes over frequencies. The colormap assigns increasingly red shades to PMUs that have a higher correlation with the PMU that has the highest FFT magnitude, to investigate the propagation of oscillation effects.

Scenario: Martin sees that for the new time period and PMU subset selected (as shown in Figure 5(II)), in (D), PMU #122 displays the highest FFT magnitude; he finds this to be consistent for the full event duration of $\sim$2 hours on scrubbing the time slider (a5). This justifies #122’s selection as the likely epicenter PMU. On juxtaposing panels (D) and (B), he also sees that (i) those PMUs topologically closer to the epicenter PMU (i.e., there are fewer hops in the network separating them) displayed higher oscillation frequency peaks over this time period, and (ii) PMUs situated at lines with the same voltage levels show similar activity.

(E) Non-Selected PMU FFTs Panel. Similarly, the FFTs of unselected PMUs are displayed in this chart. Increasingly blue shades are assigned to those PMUs that have a higher correlation with the PMU that has the highest FFT magnitude.

Scenario: Martin compares (D) and (E) to find that the highest peaks seen in (E) occur at multiple frequencies and are also of very low magnitude. He compares PMUs from (D) and (E) against flagged PMUs (a4)– PMUs from (E) are all not flagged, and are ranked very low; he, therefore, concludes that peaks from (E) can be attributed to numerical FFT noise rather than the indication of a physical event occurring.

(F) Similarity Panel. This plot uses dimensionality reduction to plot PMUs based on their similarity to each other. Specifically, t-SNE [23], short for t-distributed stochastic neighbor embedding, is an unsupervised non-linear algorithm for embedding high-dimensional data into a low-dimensional embedding space (usually two or three dimensions). We compute the Euclidean distance between PMUs based on their FFT values and plot them using t-SNE in a two-dimensional space. PMU circles placed closer together are more similar, while PMUs farther apart are less similar. To resolve overplotting, collision detection is available, which shifts PMU points so their circles do not overlap.

Scenario: Martin observes that selected PMUs are all placed in a large central cluster (Figure 5(III)), which gives a cue to the structure of the oscillation event (i.e., it likely affects points in this cluster the most), and provides the user a pool of candidate PMUs for selection. We have annotated, via dotted concentric rings, the average hop distance from the epicenter PMU. Martin also sees that the 9 flagged PMUs are located closest to the epicenter PMU #122 in this cluster. He notes that interestingly, some PMUs from the Flange and Delcino substations—one hop away from the Yearling substation—are not placed in this central cluster. On examination of these PMU locations in (B) and their FFT curves (D), he observes that these PMUs do not show high FFT peaks, and are on lower voltage lines. He infers that this is likely caused by the presence of step-down transformers within these substations between buses that connect the epicenter PMU with lines going out to these external PMUs. When the voltage is stepped down, the oscillatory behavior is damped over these lower voltage lines.

(G) Dendrogram Panel. The final plot is a novel dendrogram clustering panel, which uses a combination of $k$-means and hierarchical clustering to group selected PMUs into nodes, which are rendered in a dendrogram layout. (g1) This panel allows for any PMU (or group of PMUs) to be set as the epicenter, from which all other selected PMUs are juxtaposed. Each layer of the dendrogram corresponds to the hop distance from the selected epicenter PMU(s).

(g2) The clusters at each hop can be either automatically generated or manually specified (e.g., by selecting different $k$-values). A color swatch indicates the correlation between the cluster and the epicenter PMU(s), where darker indicates more correlation. Each cluster contains: (i) statistics about the number and type of links present as well as the number of cluster PMUs, (ii) the FFT line chart, and (iii) a box plot of the FFT value distribution of PMUs in the cluster. The clustering silhouette score [38], hop number, and total PMUs per hop are also indicated at the bottom of each dendrogram layer. This view is scrollable and zoomable, which is helpful for large dendrograms with many hops. We default to the optimal clustering for each layer based on silhouette score.

(g3) There are three types of links present in the dendrogram. (i) Self-links comprise connections between PMUs located within the same cluster, represented by a circular arc in the bottom-left corner of the node. (ii) Intra-hop links are links present between PMUs in different clusters that belong to the same hop, represented by arced and dotted lines between clusters. (iii) Inter-hop links connect PMUs in clusters belonging to different hops, represented as colored flows.

In most cases, the clusters between two hops will be fully connected, given that the grid network is itself densely connected. We therefore color the flows corresponding to the color swatch on the cluster of its origin; the flow thickness corresponds to the level of influence on downstream clusters, calculated based on the relative Manhattan distances of the cluster-wise averaged FFT data. To reduce clutter, self and intra-hop links can be toggled off, and the links turn at right angles.

(g4) To populate the dendrogram view, on enabling dendrogram PMU selection, the ’Sync’ (g5) and ’Update’ (g6) buttons are enabled. (g5) Initially, the sync button automatically includes all of the PMUs selected in (A)–(F) for clustering in the dendrogram.

Scenario: Martin moves on to tracking the nature of event propagation over the selected PMUs using the dendrogram. Initially, he loads the dendrogram for the default epicenter specified, PMU #122 (Figure 5(IV)). He observes that as the hop distance increases, the oscillation frequency drops from 2.5Hz to 0.5Hz, with the furthest selected PMUs located 5 hops away from the epicenter. He sees that all the flagged PMUs are located a maximum of 1-2 hops away, for a threshold of 0.0093. Martin deduces that the event selected is found to mainly propagate through PMUs that are connected on high voltage lines to the Delcino substation, which is located one hop away from the epicenter substation. He bases this on the ‘strong’ inter-hop flows between dendrogram nodes constituting these PMUs and the epicenter cluster; these clusters are also more closely aligned (darker colored) with the epicenter cluster. Additionally, the box plots help Martin verify that clusters are formed in a meaningful manner, as they are seen to reflect uniform distributions, such that the constituent oscillating PMUs closely influence one another.

Next, Martin updates his selection to include more PMUs located on high voltage lines at the Delcino substation, and remove unflagged PMUs. (Note: This modification still preserves the original PMU selection in previous panels.) He regenerates the dendrogram with customized clustering (3 clusters per hop). Martin can see in the new dendrogram (Figure 5(V)), with PMU #211 as the epicenter, the PMUs from the modified selection are all flagged for the set threshold. Additionally, the dominant cluster (thickest flows) in Hop 0 has PMUs located on buses for lines going from Delcino to the bus in Yearling on which the epicenter PMU is located. The FFT plots for all PMUs align and show a 2.5Hz oscillation; the box plots also show uniform distributions across all clusters. Martin can therefore verify that the system voltage oscillation propagates through PMUs on high voltage lines, over a time period of two hours.

In general, we find that 2–3 clusters per hop are generated with optimal clustering. When navigating the dendrogram view, PMUs corresponding to the view are highlighted across all other panels. In addition to event propagation tracking, the dendrogram view also enables the discovery of new grid events that might not necessarily be flagged by standard monitoring systems until they evolve to cause widespread disruptions further downstream. The flexibility afforded in potential epicenter selection drives event discovery, and constitutes the second part of our usage scenario (discovering a previously undiscovered event).

(g9) The parameters used for FFT calculation, specific to the dendrogram view, can be modified in (A). The ’Submit’ button can be used to update the dendrogram according to this (panel (A)’s parameter information reverts back to the values prior to dendrogram update). This feature allows the simultaneous analysis of event characteristics recorded by PMUs at different points in time.

Scenario: Martin decides to explore time periods outside the event duration, to check if any other significant anomalies or events preceded, succeed, or concurred with the oscillation event he analyzed. In this process, he discovers that at 12:49:00, on the same day as the analyzed oscillation event, a small anomaly is reflected in the IAm/IBm/ICm (positive current magnitude for phases A/B/C) of PMU #169 at the Sturgeon substation. Calculating FFT over 10 seconds, he finds that this is a well-damped transitory current oscillation—one that lasts for approximately 2 seconds, disappearing completely after only three cycles (as shown in Figure 5(VI)(c)).

Martin observes (from Figure 5(VI)(a) and (b)) that high overlap of FFT values is seen for immediate neighbors of the epicenter PMU #169; on further analysis, he notes that neighboring PMUs are located on low voltage lines (69kV), and are connected to step-down transformers, accounting for the highly localized oscillation (which is also symmetrical over IAm and ICm). The main frequency oscillation values for the epicenter PMU, calculated over 10 second time windows at the event time stamp, show that the system is almost completely at rest, with only a momentary sharp increase in the main frequency of the epicenter PMU to $\sim$0.7 Hz. Martin spots that the event quickly dies out as the oscillation damps, with the system returning to an “at rest” state comprised of very slow oscillations that can be filtered out using a high-pass filter. He attributes these observations to the occurrence of transient events such as “earth faults” [24] (comprising of line/generator tripping or high load fluctuation). These are known to give rise to current signals containing decaying DC components, and exhibiting such damped, transitory power oscillations.

To empirically evaluate PMU Tracker, we conducted pair analytic sessions [11] with 12 power domain experts (Table 2); four of these practitioners regularly conduct data-driven analyses on the same regional power grid networks that constitute our dataset; for these, we used the raw (non-anonymized) interface; the other participants used the anonymized version of the interface. Each pair analytic session comprised of a visualization expert well-versed with the system functionality who “drives,” while the study participant (power domain expert) freely made analysis and investigative decisions based on their own expertise and desires. Sessions lasted between 30–45 minutes (average$=37.5$, SD$=3.26$ minutes).

When conducting a session, the visualization expert’s interactions comprise largely of (i) clarifying participants’ questions regarding either the functionality of controls or the nature of data presented, (ii) obtaining clarification from the participant regarding a specific navigation action that the participant wants to perform, if received instructions are unclear. In pair analytics, freeform verbal discussion between the driver and participant is the basis for understanding the participant’s sensemaking process as well as what specific insights are uncovered during investigation. To mitigate confounds, the driver should in no way specifically prompt or leadingly question a participant; the driver may remind the participant to verbalize their thought process periodically if the participant has stopped doing so. The visualization expert’s role is therefore largely passive in the decision-making, such that maximally achievable, and objective, feedback is obtained from the participant.

We report here general feedback about the system’s analytical capabilities, as well as selected quotes from experts which help illustrate general trends in feedback.

Feedback was generally positive, with all experts particularly excited about the dendrogram panel as a new way to characterize the impact and evolution of oscillations on both an event’s epicenter PMU and its surrounding PMUs. “ [The dendrogram] in simulation-based event analysis since the flows make it easy to follow the relations between clusters.” The epicentric nature of the analysis was also appreciated: “[using] a combination of epicenters and changing the time scale will help with concurrent event identification as well,” and “coarse-grained mapping to geographical regions of interest where multiple PMUs seem to move together is afforded.”

One engineer remarked that, when used in association with the network graph contour and similarity panel, the dendrogram “allows for easy extensibility to other spectral analysis techniques like modal analysis for multiple events,” and that “event classification can be implemented based on oscillation modes, with the isolation of the most affected PMUs.” Another researcher suggested that, “line flows and line parameter estimates can be simulated on the network and dendrogram panels to map out control protocols.” These comments help demonstrate that, not only was the interface seen as useful, but it opened up new possibilities for visualizing and analyzing grid data.

The mapping of spatial PMU data with time-varying attribute values was found to be well-implemented. “It’s difficult usually, going between the map and everything else. Here the color-coding of both points [PMUs in FFT panels] and the clusters formed for the last [dendrogram] helps you hop around the network.” Some experts remarked that adding alternative or automatic selection interactions might minimize operator overhead further.“[network] PMUs could be lassoed, with a limit on the number of PMUs per substation, so we could just drag a net on the whole area. ”“[Flagging] is a neat idea to tie in with protocols in place– we get flashing notifications on screen. Maybe extending that to auto-prune everything that doesn’t matter would save time if it’s a high-stakes situation. High-quality communication, yes, but also speed in relaying to downstream operators to shift or cut transmission to protect equipment.”

One significant concern posed by a couple of experts concerned the analysis of concurrent events. “[We can] pick out anomalies and not the lower oscillation frequency noises that we might see, so we can separate out an event, fine. What do I do if I have overlapping ones? [The big drops] happen when there are multiple failure points. So we might have multiple chokes, which could stem from multiple epicenters. Not saying that we can’t find them all, just speed might be a problem since we’re doing one thing at a time. Unless we have multiple screens with different things being concurrently tested on the exact same pipeline.” “When we have [concurrent] events like that, would we see a wraparound in the dendrogram at some point? We could, especially when you start setting multiple epicenter PMUs in your initial epi-cluster. So a gradation of flows, or some sort of demarcation based on which epicenter it’s closest to would help us with stemming that effect.” We discuss this limitation in Section 7.2.

However, many experts also felt that the adaptability of the dendrogram panel to use multiple PMUs as epicenters might help address concurrent event characteristics. “I like that you can put any PMUs you want into the epicenter cluster. If I put unrelated ones in, I should know there’s another anomaly somewhere if the dendrogram flow strengths look off. So I’d have to individually look at flows for each PMU I just grouped in, one or two hops away. We don’t usually see very widespread repercussions from a single fault, it has to be a lot working together for one thing to pop up 10 links away on the network.”The functionality is very cool [for the dendrogram], particularly customizing clusters. It just needs a couple of tweaks to show different gradations in flow if you have multiple epicenters simultaneously. Maybe a toggle saying flows for the first PMU, for the second, for both, [you know] something easy to switch between a set of scenarios for potential epicenters. That would be good for discovery. I could open up multiple dendrograms below each other to compare for simultaneous faults.“

Experts also suggested integrating real-time processing functionalities in the future: “It’s pretty neat, with the flagging and the tree taking you through the network in a more intelligent manner to not overwhelm you. I think putting this with a streaming source would be great you could see everything move and freeze when you needed a closer look at something flagged.” The scalability of PMU Tracker was also found to support the streamlining of protocols followed in event analysis, though experts made a few suggestions to extend it further. “We [in most cases] only concurrently monitor and analyze 1-2 PMUs from anywhere between 5-10 substations across the grid to track an event’s propagation. [PMU Tracker] can easily support that, I’ve got 32 [on screen] here.”“I think the busiest panels are the FFTs and the dendrogram. The FFTs sort themselves out mostly, I only need to look at the x-axis position and the highest peak so I don’t mind the clutter… [additionally] I have the ranked and flagged PMUs in the settings and the similarity clusters.”“It would be good to vary say the opacity or something or group the lines into clusters sort of like what you’ve done in the dendrogram… [take the] average to scale up more.”

Based on our case study and expert reviews, we believe that PMU Tracker (i) viably supports tasks T1–T4, and (ii) handles most of the domain challenges outlined in Table 1. Here, we discuss some key takeaways and lessons learned from our design and evaluation process and discuss current limitations and future directions for tools like PMU Tracker.

We present PMU Tracker, a software tool for event localization and tracing based on power system visualization and the analysis of PMU-equipped smart grid data. We identify a set of tasks to support event analysis and design a coordinated, human-in-the-loop visualization interface to support these tasks. In particular, we utilize signal processing to localize the source of oscillation events, paired with a novel epicentric cluster dendrogram visualization to understand how the events propagate through the power grid network. We validate PMU Tracker’s design and discuss how it can be generalized for epicentric analysis in other time-varying sensor networks (such as seismogram data).