Continuous-Time Alignment and Availability-Gated Fusion of Asynchronous Multi-Rate Power System Measurements

Modern power system monitoring combines phasor measurement unit (PMU) reporting at up to 60 frames per second, SCADA scans arriving every few seconds, and event records that appear at irregular instants. These streams differ in rate by two orders of magnitude and are further corrupted by communication delay, packet loss, and missing entries. Common practice interpolates all sources onto a synchronized snapshot, which introduces stale values and interpolation distortion, while most deep sequence models assume equally spaced inputs. This paper presents CT-Align, a continuous-time fusion network that encodes each measurement channel at its native irregular timestamps with learnable exponential-decay kernels and Fourier time embeddings; aligns the sources on a small set of learnable anchors with linear-cost attention; weights every source through an availability-aware gate driven by observation density, recency, and source identity; and reconstructs bus voltage magnitudes at requested instants. On 50 held-out IEEE 39-bus disturbance episodes, drawn from a 240-episode ANDES corpus, CT-Align attains an overall RMSE of 4.72 × 10−3 p.u.: 12.4% below the strongest deep baseline retrained under the identical augmentation recipe, and 35.8% below a WLS estimator with PMU interpolation, whose transient error it halves while remaining behind it in quasi-steady windows. Deployed as one fixed checkpoint without retraining, the model loses 20.3% when channel-level SCADA loss rises from 5% to 20% and a factor of 2.6 when the SCADA source is removed altogether, a condition under which the snapshot pipeline is unobservable. On the ETTm1 transformer record, with 90% of the covariate samples removed, the seven-variable reconstruction MSE is 0.292 against 0.349 for mTAND, and a second transformer record reproduces that margin at 90%, though not at 50%. The gate weights track each source’s measured contribution to the output for the two periodic sources but not for the sparse event channel, and they are relative mixture coefficients rather than calibrated confidences. A single query takes 1.9 ms on a desktop CPU under a 100 ms cache-refresh schedule, on and between the points of the 10 Hz evaluation grid.