Traditional　methods　for　detecting　power　quality　disturbances　are　hindered　by　limited　labeled　data　and　predominantly　emphasize　classification,　often　failing　to　achieve　precise　temporal　localization.　To　address　these　issues,　this　paper　proposes　a　semi-supervised　temporal　localization　via　hierarchical　differential　attention　and　spatiotemporal　dual-threshold.　A　teacher　network,　constructed　with　differential　attention　mechanisms　and　dilated　causal　convolutions,　is　trained　on　a　small　labeled　dataset　to　capture　time-varying　features　and　generate　high-confidence　pseudo-labels　from　unlabeled　data.　To　address　the　granularity　mismatch　commonly　observed　in　conventional　pseudo-label　filtering,　a　refined　spatiotemporal　dual-threshold　mechanism　is　introduced,　which　performs　collaborative　filtering　at　both　the　time-point　and　sample　levels.　This　approach　preserves　reliable　supervision　at　high-confidence　time　points　while　suppressing　the　temporal　propagation　of　local　prediction　errors,　thereby　significantly　enhancing　the　quality　of　pseudo-labels　and　improving　localization　performance.　These　pseudo-labels　are　then　leveraged,　together　with　the　labeled　data,　to　train　a　lightweight　student　network　based　on　ShuffleNet.　Experimental　results　show　that　with　only　20　labeled　samples　per　class,　the　student　network　achieves　94.31　%　accuracy　on　simulated　data　and　98.73　%　on　real-world　data,　with　an　average　inference　time　of　just　55.12　ms.　These　findings　highlight　the　proposed　framework＇s　robustness　under　small-sample　conditions,　the　effectiveness　of　the　dual-threshold　pseudo-labeling　strategy,　and　its　practical　suitability　for　real-time　disturbance　detection　on　resource-constrained　devices.　©　2025