Alert Optimization of the PLUM Earthquake Early Warning Algorithm for the Western United States

Abstract

We determine an optimal alerting configuration for the propagation of local undamped motion (PLUM) earthquake early warning (EEW) algorithm for use by the U.S. ShakeAlert system covering California, Oregon, and Washington. All EEW systems should balance the primary goal of providing timely alerts for impactful or potentially damaging shaking while limiting alerts for shaking that is too low to be of concern (precautionary alerts). The PLUM EEW algorithm forward predicts observed ground motions to nearby sites within a defined radius without accounting for attenuation, avoiding the earthquake source parameter estimation step of most EEW algorithms. PLUM was originally developed in Japan where the alert regions and ground motions for which alerts are issued differ from those implemented by ShakeAlert. We compare predicted ground motions from PLUM to ShakeMap-reported ground motions for a set of 22 U.S. West Coast earthquakes of magnitude 4.4–7.2 and evaluate available warning times. We examine a range of prediction radii (20–100 km), thresholds used to issue an alert (alert threshold), and levels of impactful or potentially damaging shaking (target threshold). We find optimal performance when the alert threshold is close to the target threshold, although higher target ground motions benefit from somewhat lower alert thresholds to ensure timely alerts. We also find that performance, measured as the cost reduction that a user can achieve, depends on the user’s tolerance for precautionary alerts. Users with a low target threshold and high tolerance for precautionary alerts achieve optimal performance when larger prediction radii (60–100 km) are used. In contrast, users with high target thresholds and low tolerance for precautionary alerts achieve better performance for smaller prediction radii (30–60 km). Therefore, setting the PLUM prediction radius to 60 km balances the needs of many users and provides warning times of up to ∼20 s.

Incorporating Intensity Distance Attenuation Into PLUM Ground‐Motion‐Based Earthquake Early Warning in the United States: The APPLES Configuration

“…Due to the extremely limited time frame available to issue actionable EEW alerts, there will always be a trade-off between ground-motion prediction accuracy and alert timeliness. In the original PLUM approach, some of this Earth's Future 10.1029/2023EF004126 alert performance trade-off is incorporated into the choices of the fixed forward-prediction radius and the set of predefined alert regions, rather than having this trade-off be affected only by the choice of alert threshold (Cochran et al, 2022;Minson et al, 2020). PLUM generally overestimates ground motions compared to observed values; this is not surprising, as PLUM predictions are not designed to match observed ground-motion distributions.…”

“…Following the procedure in Cochran et al. (2022), for each detected event, we determine the alert qualities of the PLUM and APPLES alerts, considering different combinations of alert threshold and target threshold MMI values. We consider target thresholds ranging from MMI 3.5 to MMI 7.0 and alert thresholds ranging between MMI 2.5 to MMI 7.0, both in 0.5 MMI unit increments.…”

“…We use ShakeMap MMI distributions as the ground truth for our alert quality analysis rather than comparing to individual station observations. Station‐based analyses generally cannot comprehensively portray alert performance because the uneven station distribution can obscure Missed and Precautionary alerts as well as underestimate the sizes of late alert zones (Cochran et al., 2022; Minson et al., 2020). Performance assessments using gridded ShakeMap MMI distributions, on the other hand, allow for alert quality calculations across the entire alerted region, enabling evaluations in terms of both area and population (alerting people is a key objective for EEW systems).…”

“…We next estimate warning times at all alerted ShakeMap locations relative to the S‐wave arrival, using a fixed S‐wave velocity of 3 km/s. An analysis of the moveout distributions for different MMI exceedances determined this velocity to be appropriate for estimating exceedance times for target thresholds of MMI 3.5 and higher and yields conservative warning time estimates for target thresholds of MMI 5.0 and higher (Cochran et al., 2022). While the ShakeAlert Testing Platform simulates real‐time data streams by sending station data to the EEW algorithms in one‐second packets, it does not include data telemetry latencies or additional latencies from alert delivery, both of which can range from <1 s to several seconds or more (McBride et al., 2023; Patel & Allen, 2022; Stubailo et al., 2021).…”

“…Secondly, the warning times for PLUM alerts are determined primarily by the size of the prediction radius, which is a fixed value for all ground motions. Thus, simply shrinking the prediction radius to try to reduce over‐alerting will shorten warning times everywhere and potentially cause more late and missed alerts (Cochran et al., 2022). For large magnitude earthquakes that produce high near‐source ground‐motion amplitudes and large footprints of significant shaking, this fixed prediction radius also limits the warning times available at locations outside of the immediate source area (Kodera, 2019; Meier et al., 2020).…”