5 results about "Temporal similarity" patented technology

The invention discloses a reservoir group long series scheduling process multi-objective optimization method based on space-time similarity dimensionality reduction, and belongs to the technical field of reservoir scheduling of hydraulic engineering. Performing space-time multi-scale dimension reduction on the scheduling process to obtain a dimension-reduced scheduling process; according to the scheduling process after dimension reduction, based on a multi-objective optimization scheduling algorithm, similarity judgment is carried out on the offspring individuals; directly reusing the parent scheduling process for the individuals which are determined to be similar to obtain corresponding objective function values, and scheduling the reservoir group scheduling model for the individuals which are determined to be dissimilar to obtain corresponding objective function values; and updating the non-dominated solution set according to the obtained objective function value, and generating an optimized reservoir group length series scheduling process. According to the method, the optimized reservoir group length series scheduling process can be quickly generated on the premise of ensuring the quality of the solution set.

This invention discloses a method and system for spatiotemporal prediction of charging station load based on multi-source matrix fusion. The method includes: collecting historical load data and geographical coordinates of charging stations, performing standardization and sample partitioning; constructing in parallel a prior adjacency matrix based on geographical distance, a temporal similarity adjacency matrix based on dynamic time warping distance, and a dynamic adaptive adjacency matrix based on learnable node embedding; adaptively weighting and fusing the three matrices through trainable scalar parameters to generate a fused adjacency matrix; and constructing and utilizing a spatiotemporal graph convolutional network for load prediction based on the fused adjacency matrix as the spatial relationship basis. This invention comprehensively improves the accuracy and reliability of spatiotemporal prediction of charging station group load by fusing complementary spatial relationships from multiple sources, enhancing the robustness of temporal similarity measurement through dynamic time warping, and combining data-driven adaptive learning.

Ultrasonic image processing method and ultrasonic imaging device The present invention relates to an ultrasonic image processing method comprising at least: - decomposition (204) into singular values ​​of each image of at least one initial sequence of ultrasonic images, calculating spatial eigenvectors and / or temporal eigenvectors of each of the images; - partitioning (206) of said spatial eigenvectors and / or said temporal eigenvectors, distributing said spatial eigenvectors and / or said temporal eigenvectors into several groups of spatial eigenvectors and / or several groups of temporal eigenvectors on the basis respectively of spatial and / or temporal similarities; - calculation (208) of at least one final sequence of images by concatenation of said spatial eigenvectors and / or said temporal eigenvectors of one of the groups of vectors. Figure for the summary: Fig. 2

A temporal sequence of multi-modal signals is generated from a feature probe signal, a relation probe signal, and attribute probe signal, and multi-modal signals are selected from the temporal sequence of multi-modal signals. The selected multi-modal signals are compared to a model multi-modal embedding space cluster to generate the multi-modal temporal similarity measures. The multi-modal temporal similarity measures are compared to a model similarity measure boundary to generate object recognition verification data associated with an object classification.

The invention discloses an irregular multivariable time sequence prediction method and device based on patch block pre-alignment, which are applied to an irregular multivariable time sequence prediction model, and the method comprises the following steps: a patch block aligner performs efficient mapping on an irregular time sequence to obtain an aligned regular time sequence; carrying out layer-by-layer aggregation on the fine-grained patch blocks, the medium-grained patch blocks and the coarse-grained patch blocks by a multi-scale time encoder to construct multi-scale time embedding features; the group mixer carries out fusion of time period features and channel features on the multi-scale time embedding features to obtain global representation; and the prediction output module processes the global representation to obtain an irregular time sequence prediction result. According to the method provided by the invention, an efficient patch block type pre-alignment mechanism is introduced, the multi-scale time similarity is reserved in the modeling process, the problem that a traditional prediction algorithm is difficult to adapt to irregular sampling data is solved, the limitations that an existing irregular prediction model is unstable in performance, too large in calculation overhead and the like are overcome, and the prediction precision is remarkably improved.