Dynamic correlation modeling and algorithm optimization of athletes' training load and performance based on deep time series network

Dynamic correlation modeling between training load and sports performance is central to scientific training. The traditional Banister impulse-response (IR) model, constrained by a linear assumption and one-dimensional load input, cannot characterize multi-time-scale coupling or individual heterogeneity. This paper proposes a multi-scale temporal attention network (MS-TANet) with three collaborative modules: a multi-scale dilated causal convolution (MS-DCC) module that extracts day-, week- and month-scale load features in parallel and fuses them by learnable gating; a sparse decay self-attention (SD-SA) module that embeds a learnable exponential decay bias to identify training-effect lag windows, reducing asymptotic complexity to \(O(L(logL+k{D}ₖ))\) via top-k sparsification with locality-sensitive hashing; and an individual embedding and feature modulation (IE-FM) module that conditions temporal features on static attributes and monthly-refreshed physiological covariates. A physics-constrained regularization loss and a three-stage curriculum strategy embed physiological priors and mitigate small-sample overfitting. Unlike TFT, which treats attention as purely data-driven, MS-TANet elevates the physiological time constant to an interpretable learnable parameter; unlike physics-informed networks with hard residual constraints, it embeds fatigue decay and adaptation boundedness as soft regularization, forming a “local–long-range–individual” complementary structure. On three datasets, MS-TANet attains R 2 of 0.889, 0.871 and 0.853—4.0 percentage points (4.8% relative) above the best baseline TFT and 21.7 points (33.4% relative) above Banister IR. Shapley decomposition attributes 26.7%, 23.4% and 11.2% of the gain to SD-SA, IE-FM and their interaction. The learned half-lives (16.9–21.0 days) match physiological ranges, though this partly reflects the built-in decay bias rather than a free discovery. Attention heatmaps pass window-removal, perturbation and randomization checks, and 95% prediction intervals reach 93.7% empirical coverage.