Cloud-edge collaborative intelligent scheduling system

Abstract

The application relates to the technical field of industrial artificial intelligence, and particularly relates to a cloud-edge collaborative intelligent scheduling system. In view of the problems of insufficient description of dynamic topology changes, mutual interference of multi-task prediction and uneven task distribution in existing centralized scheduling, which leads to the decline of real-time performance and throughput, a cloud-edge collaborative scheme is provided: the cloud end extracts semantic features of industrial system data, fuses graph space-time and generates link priori, and then divides the tasks according to the total running time and dynamically distributes them to the edge side; the edge side completes multi-task time sequence prediction in the task-specific low-dimensional subspace, the optimization side further constructs the prediction result into a probability scenario and embeds it into the decision-making process, and outputs a probability weighted control instruction. The scheme can improve the prediction stability, scheduling robustness and resource utilization, and improve the system throughput, real-time performance and control economy.

Description

Technical Field

[0001] This application relates to the field of industrial artificial intelligence technology, and in particular to a cloud-edge collaborative intelligent scheduling system. Background Technology

[0002] The scheduling system in the Industrial Internet typically adopts a centralized architecture, combining time-series prediction, reinforcement learning scheduling, and topology analysis to achieve equipment operation and task allocation management.

[0003] However, existing solutions mostly rely on centralized cloud computing, which is difficult to adapt to scenarios with multiple sources of uncertainty and multiple concurrent tasks; static topology analysis is difficult to characterize topology evolution and uncertainty propagation, shared prediction models are prone to inter-task state interference, and averaged task distribution methods are prone to load imbalance, throughput reduction and insufficient real-time performance.

[0004] Therefore, how to improve scheduling robustness, prediction stability and system throughput under conditions of multi-source uncertainty and multi-task concurrency has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a cloud-edge collaborative intelligent scheduling system to address the aforementioned technical problems. This system targets scheduling scenarios in the Industrial Internet characterized by multi-source uncertainty and multi-task concurrency. By constructing a cloud-edge collaborative task processing and scheduling framework, it integrates industrial system data analysis, scheduling task generation and distribution, multi-task predictive processing, and scheduling decision optimization, thereby improving the adaptability of the scheduling process to uncertainties in complex industrial scenarios, the stability of prediction results, and overall task processing efficiency.

[0006] Firstly, this application provides a cloud-edge collaborative intelligent scheduling system, the system comprising:

[0007] The cloud processing unit is used to extract semantic features and fuse graph spatiotemporal structures from industrial system data, and generate prior features for link prediction.

[0008] The distribution unit is used to perform non-uniform partitioning and encapsulation of the scheduling task containing the prior features based on the total running time model to generate prediction tasks, and to dynamically instantiate and distribute the prediction tasks to the edge gateway.

[0009] The prediction unit is used to assign a mapping vector to each prediction task to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state, and to solve the shared output layer weights based on the projected state to obtain the multi-task time series prediction results.

[0010] The optimization unit is used to generate a probabilistic scenario based on the multi-task time-series prediction results, embed the probabilistic scenario into a Markov decision process, and output scheduling action instructions using a policy optimization algorithm.

[0011] In one possible implementation, the cloud processing unit includes a semantic extraction module, a graph structure modeling module, and a spatiotemporal evolution analysis module, wherein:

[0012] The semantic extraction module is used to extract semantic features from the node and edge attributes in the industrial system data through the large language model, and generate semantic embedding vectors.

[0013] The graph structure modeling module is used to construct a weighted graph structure based on the semantic embedding vector, and to perform spatial aggregation on the weighted graph structure through a graph convolutional network to generate spatial features.

[0014] The spatiotemporal evolution analysis module is used to perform temporal evolution analysis on the spatial features and generate prior features for link prediction.

[0015] In one possible implementation, the graph structure modeling module is specifically used for:

[0016] Based on the semantic embedding vector and the device connection relationships and operating status sequences in the industrial system data, the weighted graph structure is constructed;

[0017] The spatial features are generated by spatially aggregating the weighted graph structure using a graph convolutional network.

[0018] The spatial features are processed through a robust identification mechanism, and the links with high uncertainty are downweighted to filter out noise interference.

[0019] In one possible implementation, the spatiotemporal evolution analysis module is specifically used for:

[0020] The spatial features after robust identification processing are analyzed over time using a time-series modeling algorithm to obtain the prior features for link prediction.

[0021] In one possible implementation, the time-series modeling algorithm includes a bidirectional long short-term memory network and a Transformer architecture. The bidirectional long short-term memory network is used to extract local short-range fluctuation features in the spatial features, and the Transformer architecture is used to capture global long-range dependencies and periodic evolution patterns in the spatial features.

[0022] In one possible implementation, the distribution unit includes a task partitioning module, specifically used for:

[0023] Construct a total runtime model that includes startup overhead, communication latency, and computation time;

[0024] Based on the total runtime model, the scheduling task containing the prior features is dynamically and non-uniformly partitioned and encapsulated to obtain the predicted task, so that the first-executed shards can obtain more resources to mask the startup delay, and the later-executed shards are dynamically adjusted to eliminate tail-end idleness.

[0025] In one possible implementation, the distribution unit further includes a resource management module, specifically used for:

[0026] Based on the non-uniform partitioning result, two-layer dynamic instantiation is performed to establish a distributed computing instance set, and the prediction task is sent to the edge gateway.

[0027] The first-level instantiation allocates basic container groups based on the divided task fragments, while the second-level instantiation dynamically expands and shrinks based on the length of the real-time task queue.

[0028] In one possible implementation, the prediction unit includes a subspace mapping module, specifically used for:

[0029] For each prediction task, a binary mapping vector with a specific proportion of zero elements is generated, and the binary mapping vector is subjected to a Hadamard product operation with the shared reservoir state to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state.

[0030] In one possible implementation, the prediction unit further includes an output solution module, specifically used for:

[0031] Based on the projection state, ridge regression is used to solve for the shared output layer weights to obtain the time-series prediction results of the prediction task.

[0032] The multi-task time series prediction results are obtained by summarizing multiple time series prediction results.

[0033] In one possible implementation, the optimization unit is specifically used for;

[0034] Based on the multi-task time-series prediction results, multiple probabilistic scenarios are generated using Latin hypercube sampling, and a representative scenario set is obtained through synchronous back-substitution reduction.

[0035] The probability of each scenario in the representative scenario set is used as a weight to embed into the reward function of the Markov decision process, and the scenario variables are extended to the state space to construct a Markov decision process with probability weighting.

[0036] A shearing substitution mechanism with near-end policy optimization is used to constrain the policy update step size, and the policy is updated for the probability-weighted Markov decision process, outputting the scheduling action instruction.

[0037] This application provides a cloud-edge collaborative intelligent scheduling system. Through a cloud processing unit, it performs semantic feature extraction based on a large language model, graph spatiotemporal structure fusion, and link prediction prior feature generation on industrial system data. This improves the system's ability to represent the state, topological relationships, and uncertainty propagation, thereby enhancing the robustness of scheduling decisions. The distribution unit performs non-uniform partitioning and encapsulation of scheduling tasks containing prior features based on a total runtime model, and, in conjunction with two layers of dynamic instantiation and distribution of prediction tasks, improves the rationality of task allocation and edge-side resource adaptation, thereby improving load balancing, system throughput, and real-time response capabilities. Through pre-... The measurement unit assigns a binary mapping vector to each prediction task and projects it onto a low-dimensional task-specific subspace. Based on the state of the projected subspace, it solves for the shared output layer weights, which can reduce state interference between multiple tasks and improve the stability of temporal prediction, thus providing more reliable prediction results for subsequent scheduling optimization. The optimization unit generates probabilistic scenarios based on the temporal prediction results of multiple tasks and reduces them. The scenario probabilities are embedded into a Markov decision process and a near-end policy optimization is used for policy updates. This enables the output of action instructions that are more consistent with the probability distribution characteristics in uncertain scenarios, thereby improving the overall scheduling performance of the cloud-edge collaborative intelligent scheduling system in complex industrial environments. Attached Figure Description

[0038] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0039] Figure 1 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 1 ;

[0040] Figure 2 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 2 ;

[0041] Figure 3 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 3 ;

[0042] Figure 4 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 4 ;

[0043] Figure 5 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 5 .

[0044] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0045] Industrial Internet scheduling technology is mainly applied in scenarios such as intelligent manufacturing, energy management, and equipment joint control. It is used to predict, distribute, and control under conditions of equipment status fluctuations, load changes, and multi-task concurrency. Related systems typically rely on a cloud-based centralized computing architecture to aggregate and analyze industrial system data and issue scheduling commands to field edge nodes or executing equipment.

[0046] Existing solutions typically employ a centralized scheduling approach, deploying topology analysis, time-series prediction, and reinforcement learning decision-making centrally in the cloud. They receive data uploaded from the industrial site via a unified task queue and then distribute tasks and output control based on pre-defined rules. While this approach can achieve basic equipment operation management, its topology modeling often remains at the level of static correlation analysis. It struggles to reflect the dynamic evolution of link relationships in industrial networks as operating conditions change, and it also fails to characterize the propagation of uncertainty between different devices and links.

[0047] Meanwhile, in multi-task concurrent scenarios, existing prediction models often use a shared state space to process different types of tasks, which can easily lead to mutual interference between task features. This results in significant distortion of prediction results when load fluctuates, renewable energy output changes, or equipment coupling increases. In addition, the averaged task distribution method lacks awareness of differences in startup overhead, communication latency, and computation time, which can easily lead to some edge nodes being overloaded while others are idle. This results in decreased system throughput, insufficient real-time performance, and makes scheduling strategies more conservative, making it difficult to balance robustness and economy.

[0048] In view of this, how to improve prediction stability, scheduling robustness, and overall throughput under conditions of multi-source uncertainty and multi-task concurrency has become an urgent technical problem to be solved. To address the above problems, this application provides a cloud-edge collaborative intelligent scheduling system. The system uses a cloud processing unit to extract semantic features and fuse graph spatiotemporal structures from industrial system data based on a large language model, and generates prior features for link prediction. Then, the distribution unit performs non-uniform partitioning based on the total runtime model and distributes prediction tasks in combination with two-layer dynamic instantiation. Subsequently, the prediction unit completes task-specific subspace projection and multi-task temporal prediction. Finally, the optimization unit embeds the probabilistic scenario into a Markov decision process and uses a near-end strategy to optimize the output scheduling action instructions, thereby achieving collaborative cooperation between cloud-based macro-analysis and edge-side real-time scheduling.

[0049] Figure 1A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 1 .like Figure 1 As shown, the cloud-edge collaborative intelligent scheduling system includes:

[0050] The cloud processing unit is used to extract semantic features and fuse graph spatiotemporal structures from industrial system data, and generate prior features for link prediction.

[0051] The distribution unit is used to perform non-uniform partitioning and encapsulation of scheduling tasks containing prior features based on the total running time model to generate prediction tasks, and to dynamically instantiate and distribute the prediction tasks to the edge gateway.

[0052] The prediction unit is used to assign a mapping vector to each prediction task to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state, and solve the shared output layer weights based on the projected state to obtain the multi-task time series prediction results.

[0053] The optimization unit is used to generate probabilistic scenarios based on multi-task time-series prediction results, embed the probabilistic scenarios into Markov decision processes, and output scheduling action instructions using policy optimization algorithms.

[0054] The cloud processing unit, deployed in the cloud, acts as the global decision-making brain. It extracts semantic features from industrial system data using a large language model and fuses the spatiotemporal structure of the graph. It also generates prior features for link prediction through a robust recognition mechanism. The cloud receives industrial system data, which includes topological data such as device nodes and edge connections, as well as attribute data such as device models and operating parameters. After link prediction and feature extraction, macroscopic prior features containing topological weights and evolutionary trends are formed. These features are responsible for macroscopic technology evolution prediction, cross-node semantic feature extraction, and high-throughput task distribution, primarily handling long-cycle, large-scale topological evolution and resource coordination. The distribution unit acts as an intermediary bridge between the cloud and the edge. It performs non-uniform partitioning and encapsulation of scheduling tasks containing prior features based on a total runtime model to generate prediction tasks. It then coordinates with two layers of dynamic instantiation to distribute prediction tasks to the edge gateway. This unit receives macroscopic instructions from the cloud and performs fine-grained decomposition based on the heterogeneous computing power and network status of the edge, ensuring low latency and high reliability during the distribution process.

[0055] The prediction and optimization units are deployed at the edge. The prediction unit, acting as a local execution core, allocates a binary mapping vector to each prediction task, projecting it onto a low-dimensional task-specific subspace. Based on the projected subspace state, it solves for the shared output layer weights, obtaining multi-task temporal prediction results. Upon receiving instructions at the edge, this unit concurrently executes subspace mapping predictions, transforming the information flow from feature data into multi-dimensional temporal prediction results. The optimization unit generates probabilistic scenarios based on the multi-task temporal prediction results and reduces them. It embeds the scenario probabilities into a Markov decision process, uses proximal policy optimization for policy updates, and outputs scheduling action instructions. By embedding the probabilistic scenarios generated from the prediction results into a Markov decision process and combining it with proximal policy optimization to output scheduling action instructions, the optimization unit balances robustness and economy, creating a collaborative closed loop between cloud-based macro-analysis and rapid edge execution.

[0056] Specifically, the cloud processing unit extracts semantic relationships from industrial system data using a large language model and jointly models them with the graph spatiotemporal structure, thereby more accurately representing dynamic changes and uncertainty propagation in the links. The distribution unit performs non-uniform partitioning based on the total runtime model, taking into account startup overhead, communication latency, and computational load, enabling prediction tasks to better match the processing capabilities of different edge gateways, thus improving overall throughput and real-time performance. The prediction unit uses task-specific low-dimensional subspaces to suppress feature interference between multiple tasks and utilizes shared output layer weights to maintain transferable information between tasks, thus improving prediction stability under load fluctuations and device coupling scenarios.

[0057] Figure 2 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 2 .like Figure 2 As shown, the cloud processing unit within the cloud-edge collaborative intelligent scheduling system includes: a semantic extraction module, a graph structure modeling module, and a spatiotemporal evolution analysis module, wherein:

[0058] The semantic extraction module is used to extract semantic features from the node and edge attributes in industrial system data through a large language model, and generate semantic embedding vectors.

[0059] The graph structure modeling module is used to construct a weighted graph structure based on semantic embedding vectors, and to perform spatial aggregation on the weighted graph structure through a graph convolutional network to generate spatial features.

[0060] The spatiotemporal evolution analysis module is used to perform temporal evolution analysis on spatial features and generate prior features for link prediction.

[0061] Specifically, the semantic extraction module performs unified semantic encoding on equipment type, operating condition description, connection relationships, and operational events, transforming heterogeneous industrial data into comparable semantic embedding vectors, thereby reducing the expression differences between data from different sources. The graph structure modeling module forms a weighted graph structure that characterizes the strength of node associations based on the semantic embedding vectors, and uses a graph convolutional network to aggregate neighborhood information, enabling a more accurate depiction of coupling relationships between equipment and the scope of link influence. The spatiotemporal evolution analysis module further models spatial features in the temporal dimension, identifying the dynamic evolution of link relationships with load fluctuations, equipment status changes, and task concurrency, thus outputting more forward-looking link prediction prior features. Therefore, the cloud processing unit can simultaneously characterize semantic information, spatial relationships, and temporal changes in industrial networks, improving the stability and accuracy of link prediction and providing more reliable prior support for subsequent scheduling task distribution and edge prediction.

[0062] Furthermore, the graph structure modeling module is specifically used for:

[0063] Based on semantic embedding vectors and the device connection relationships and operating status sequences in industrial system data, a weighted graph structure is constructed. The weighted graph structure is spatially aggregated through a graph convolutional network to generate spatial features. The spatial features are processed through a robust recognition mechanism to reduce the weight of links with high uncertainty in order to filter out noise interference.

[0064] Specifically, given an industrial topology graph structure G=(V,E), where V is the set of nodes and E is the set of edges, the node feature matrix is... The edge feature matrix is:

[0065] The graph structure modeling module first transforms high-dimensional edge features into scalar weights required for graph convolution using an edge weight mapping formula: The calculation formula is:

[0066] .

[0067] in, Let be the feature vector of the edge between nodes i and j. and For mapping parameters, The edge weight mapping value ranges from (0,1) and represents the semantic confidence of the link connection.

[0068] Based on this edge weight mapping value, the graph structure modeling module performs a weighted graph convolution operation to aggregate spatial information. The calculation formula is as follows:

[0069] .

[0070] in, For nodes The first convolutional layer output, For the set of neighboring nodes, This is the weight matrix. For bias. This formula is obtained through... It achieves differentiated information aggregation based on semantic confidence, thereby capturing static physical dependencies and potential associations between device nodes.

[0071] To suppress noise links and anomalous fluctuations prevalent in industrial data, the module introduces an RSCN robust identification mechanism. This mechanism calculates the fuzzy membership degree of samples. This reflects the reliability of a sample belonging to a real link, with a value range of [0,1]. Global parameters are then calculated based on this membership degree using weighted averages. :

[0072] .

[0073] in, The true labels for the links are defined. This derivation uses membership weighting to ensure that high-confidence samples contribute significantly to parameter solving, while suppressing low-confidence noisy links, thereby eliminating Q-value estimation bias and improving the robustness of link prediction.

[0074] Regarding parameter configuration, the output dimension of the graph convolutional network in the graph structure modeling module is set to 256, and constrained to integer powers of 2 to adapt to tensor core acceleration; node feature dimensions... The dimension is 16, determined by PCA dimensionality reduction from the original sensor parameters; the edge feature dimension is... The hidden layer dimension of the large language model is 256, encompassing textual semantics and historical statistical features extracted from the large language model; the hidden layer dimension of the large language model is 2048, limited by cloud memory and inference latency constraints. The coordinated configuration of these parameters ensures that the graph structure modeling module achieves efficient and robust spatial feature extraction in industrial topology environments.

[0075] By combining semantic embedding vectors with device connection relationships and operating state sequences for weighted graph structure construction, the physical connections, semantic associations, and operating condition change characteristics between devices can be simultaneously represented, thus freeing the graph structure from being limited to static topological descriptions. Spatial aggregation of the weighted graph structure using graph convolutional networks can extract the association propagation patterns and local coupling features between multiple devices, enabling spatial features to reflect the linkage state in complex industrial networks. Furthermore, robust identification mechanisms are used to process spatial features and reduce the weight of links with high uncertainty, suppressing the amplification effect of abnormal fluctuations, noisy data, and unstable connections on feature propagation. This improves the reliability and stability of spatial modeling results, providing more accurate input for subsequent spatiotemporal evolution analysis and link prediction prior feature generation. Therefore, it is beneficial for improving prediction robustness and scheduling accuracy under multi-source uncertain environments.

[0076] Furthermore, the spatiotemporal evolution analysis module is specifically used for:

[0077] By performing temporal evolution analysis on the spatial features after robust identification processing using a time-series modeling algorithm, prior features for link prediction are obtained.

[0078] The spatial features, after robust identification processing, have mitigated the impact of highly uncertain links and noise disturbances. When input into a time-series modeling algorithm, they can depict the continuous changes in device connection strength, load coupling relationships, and operating states along the time dimension, thereby extracting trend information on the transformation of link relationships from static association to dynamic evolution. This prior feature provides a more stable temporal basis for subsequent prediction tasks, enabling link prediction to reflect not only the current spatial structure but also the propagation path and evolution direction under fluctuating operating conditions. This improves the accuracy and consistency of prediction results under multi-source uncertainty conditions, thus facilitating earlier detection of potential link changes in subsequent task distribution and edge scheduling, balancing system robustness and real-time response capabilities.

[0079] The temporal modeling algorithm includes a bidirectional long short-term memory network and a Transformer architecture. The bidirectional long short-term memory network is used to extract local short-range fluctuation features in spatial features, while the Transformer architecture is used to capture global long-range dependencies and periodic evolution patterns in spatial features.

[0080] Bidirectional Long Short-Term Memory (LSTM) networks jointly encode the forward and reverse changes of spatial features at adjacent time points, enabling the identification of sudden device state changes, load fluctuations, and short-term link disturbances, thus fully preserving local fluctuation information. The Transformer architecture, based on an attention mechanism, models the correlation of features across time periods, further characterizing the long-range dependencies and periodic evolution trends under multi-device coupling relationships, thereby simultaneously considering short-term sensitivity and long-term stability. The synergistic effect of these two technologies reduces local distortions or omissions of global information caused by relying solely on a single time-series model, making the prior features for link prediction more complete and robust. This improves the accuracy and consistency of prediction results in multi-task concurrent scenarios, thus benefiting task distribution, edge execution, and control decisions in subsequent cloud-edge collaborative scheduling.

[0081] Figure 3 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 3 .like Figure 3 As shown, the distribution unit within the cloud-edge collaborative intelligent scheduling system includes a task partitioning module and a resource management module, wherein:

[0082] The task partitioning module is used to construct a total runtime model that includes startup overhead, communication latency, and computation time. Based on the total runtime model, the scheduled tasks containing prior features are dynamically and non-uniformly partitioned and encapsulated to obtain predicted tasks, so that the first-executed fragments can obtain more resources to mask the startup latency, and the later-executed fragments are dynamically adjusted to eliminate tail idleness.

[0083] The resource management module is used to perform two-layer dynamic instantiation based on the non-uniform partitioning results, establish a set of distributed computing instances, and distribute prediction tasks to the edge gateway. The first layer of instantiation allocates basic container groups according to the partitioned task shards, and the second layer of instantiation dynamically expands and shrinks according to the length of the real-time task queue.

[0084] In the specific implementation process, the task division module first establishes a total runtime model. The total runtime is precisely modeled as the sum of the startup time and the runtime, and the calculation formula is:

[0085] .

[0086] in, This refers to startup time, including container initialization and dependency loading. Pure runtime, Let be the number of instances in the second layer to be optimized, and are integers and satisfy , This represents the total number of concurrent tasks.

[0087] Startup time The segmented derivation takes into account pipeline parallelism and scheduling latency during instance startup, and the calculation formula is:

[0088] .

[0089] in, This represents the average startup time, which can range from (5, 30) seconds and is affected by the container image size. This represents the scheduling time for a single task, ranging from (0.1 to 2) seconds, and is affected by network latency. The scheduling overhead increases linearly with the number of instances.

[0090] runtime The derivation is based on the assumption of uniform task distribution, and the calculation formula is:

[0091] .

[0092] in, The average runtime is , which can range from (10, 600) seconds and is determined by the complexity of the prediction and optimization algorithms; ⌈x⌉ indicates rounding up x to ensure that all tasks are executed.

[0093] To eliminate idle tasks at the tail end, the task partitioning module introduces a dynamic adjustment formula. To resolve the issue of the last batch of instances being idle due to rounding operations, the corresponding calculation formula is:

[0094] .

[0095] in, For the first The number of tasks allocated in a batch. This derivation ensures that the remaining tasks are evenly distributed among the remaining instances, minimizing the maximum completion time.

[0096] The non-uniform task partitioning algorithm described above allocates more resources to the first-executed shards to mask startup latency, while the later-executed shards dynamically adjust to eliminate tail-end idleness. This accurately models the total runtime as the sum of startup and runtime, and further eliminates tail-end idleness by optimizing the number of second-layer instances. This effectively solves the cloud scheduling throughput bottleneck caused by multi-task concurrency at the edge, significantly improving the overall efficiency and resource utilization of task scheduling.

[0097] The resource management module employs a centralized resource manager for macro-level resource pool partitioning and priority sorting, ensuring global fairness and preventing deadlocks. Under this centralized resource manager, a distributed resource manager is introduced for two-layer dynamic instantiation. The first layer allocates basic container groups based on regional computing power, while the second layer dynamically scales up or down based on the length of the real-time task queue.

[0098] For heterogeneous tasks, the resource management module employs a core second allocation strategy to ensure load balancing across distributed nodes for tasks of varying complexity. The formula for heterogeneous task core second allocation is as follows: Used to determine the core quota for task type i:

[0099] .

[0100] in, For the first The class task weight reflects the computational complexity of a single task and is obtained by normalizing the average historical running time. This represents the number of tasks of this type; Total_Cores represents the total number of available cores in the system. This formula is based on a weighted allocation to ensure that tasks with high computational demands receive matching computing resources.

[0101] Through the resource management module, a two-layer dynamic instantiation architecture is used to achieve an organic combination of rapid deployment and real-time elastic scaling of basic container groups. At the same time, based on the core second-period allocation strategy, the load balancing of heterogeneous tasks among distributed nodes is effectively guaranteed, avoiding resource waste and task backlog, thereby significantly improving system throughput and scheduling fairness in multi-task concurrent scenarios at the edge.

[0102] Figure 4 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 4 .like Figure 4 As shown, the prediction unit in the cloud-edge collaborative intelligent scheduling system includes a subspace mapping module and an output solution module, wherein:

[0103] The subspace mapping module is used to generate a binary mapping vector with a specific proportion of zero elements for each prediction task, and to perform a Hadamard product operation between the binary mapping vector and the shared reservoir state to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state.

[0104] The output solution module is used to solve the shared output layer weights based on the projected state using ridge regression to obtain the temporal prediction results of the prediction task; it also aggregates multiple temporal prediction results to obtain multi-task temporal prediction results.

[0105] Specifically, the subspace mapping module first assigns a unique binary mapping vector to each prediction task. Let the shared reservoir state be... ,Task The binary mapping vector is The elements are either 0 or 1. The subspace projection formula achieves state isolation through element-wise multiplication:

[0106] .

[0107] in, For Hadama accumulation. The positions corresponding to elements with 0 are masked, while the positions corresponding to elements with 1 retain their original state, thus achieving selective filtering of dimensions.

[0108] The derivation of the expected overlap after mapping is as follows: Let p be the proportion of 0 elements in the mapping vector, and the probability that two different task mapping vectors are both 1 in the same dimension is (1−p)². The overlap of the original unmapped state is considered to be 1. Then the ratio of the expected overlap after mapping to the original overlap is:

[0109] .

[0110] This derivation proves that by introducing the proportion of zero elements... The state overlap between tasks decreases linearly from 1 to This eliminates state aliasing in a probabilistic sense.

[0111] To ensure that the projection operation does not disrupt the echo state characteristics of the reservoir, the subspace mapping module introduces norm inequalities:

[0112] .

[0113] This inequality shows that the lower bound of the energy of the projected state is the same as that of the original state. Times, as long as If the value is set reasonably, the short-term memory capacity of the reservoir can be maintained.

[0114] Regarding parameter configuration, the proportion of zero elements in the mapping vector. The value should be between 2% and 10%; too small a value will prevent decoupling, while too large a value will lead to information loss and leakage of integral parameters. The memory decay rate for reservoir state updates is controlled, with a typical value of 0.3; the number of steps in the washing stage is set to 100, which is used to initialize the reservoir state and eliminate the influence of zero initial value.

[0115] The beneficial effects of the subspace mapping module are as follows: by using the Hadamard projection operation, while preserving the rich dynamic characteristics of the reservoir, it physically blocks the mutual interference of different task state updates, fundamentally eliminates the aliasing of multi-task states, and provides a guarantee for the numerical stability and convergence of long-term edge prediction.

[0116] The output solution module uses ridge regression to solve for the shared output layer weights after projection, avoiding the edge computing overhead caused by complex backpropagation. The ridge regression formula is used to calculate the output weights, avoiding matrix singularity issues.

[0117] .

[0118] in, The regularization coefficient is . As a unit array, The target output is [output term]. The introduction of the regularization term ensures that a stable analytical solution can still be obtained even when multi-task concurrency leads to an ill-conditioned covariance matrix.

[0119] Regarding parameter configuration, the ridge regression regularization coefficient Values This is used to prevent overfitting and constrain the norm of the output weights. Reservoir size Typically, the number of tasks is taken as 50-100 times to ensure sufficient dimensions for subspace partitioning.

[0120] After obtaining the time series prediction results of each prediction task, the output solution module summarizes them to form the final multi-task time series prediction results.

[0121] The beneficial effects of the output solution module are as follows: by solving the shared output layer weights through ridge regression, the edge computing overhead caused by complex backpropagation is avoided. At the same time, the introduction of regularization terms ensures that a stable analytical solution can still be obtained when the covariance matrix is ​​ill-conditioned due to multi-task concurrency, thereby guaranteeing the computational efficiency and numerical stability of multi-task time series prediction at the edge.

[0122] Figure 5 A schematic diagram of the structure of a cloud-edge collaborative intelligent scheduling system provided in this application embodiment. Figure 5 .like Figure 5 As shown, the optimization unit within the cloud-edge collaborative intelligent scheduling system is specifically used for:

[0123] Based on the results of multi-task time series prediction, multiple probabilistic scenarios are generated using Latin hypercube sampling, and a representative scenario set is obtained through synchronous back-substitution reduction.

[0124] The probability of each scenario in the representative scenario set is embedded as a weight into the reward function of the Markov decision process, and the scenario variables are extended to the state space to construct a Markov decision process with probability weighting.

[0125] A shearing substitution mechanism with proximal policy optimization is used to constrain the policy update step size, and the policy is updated for a Markov decision process with probability weighting, outputting scheduling action instructions.

[0126] Specifically, the optimization unit first uses a GRU network to probabilistically predict renewable energy and load, and then generates a large number of initial scenarios using Latin hypercube sampling. To reduce the computational dimensionality of subsequent optimizations, a synchronous back-substitution reduction algorithm is used to eliminate redundant and similar scenarios based on probabilistic distance, retaining representative extreme values ​​and typical scenario sets. Let the scenario set within the scheduling period be denoted as . Scene The probability of occurrence is ,satisfy .

[0127] The reduced probability scenario is embedded into the state and reward functions of a Markov decision process, enabling the agent to perceive the uncertainty distribution during training. The overall cost objective function is then applied. Defined as the weighted sum of operating costs and uncertainty penalty costs:

[0128] .

[0129] in, Operating costs (including power generation costs and electricity purchase costs). The cost of uncertainty penalties (including load shedding penalties and wind and solar curtailment penalties).

[0130] Scenario-based probability-weighted reward function Incorporating scenario probabilities into immediate rewards guides the agent to focus on the benefits of high-probability scenarios:

[0131] .

[0132] in, For the scene The probability, For operating costs, To incur penalties, The penalty discount factor is introduced. The introduction of probability multipliers makes the policy more constrained by high-probability anomaly scenarios.

[0133] A cut-and-substitution optimization mechanism employing a proximal strategy is used to output multi-scenario probability-weighted actions. GAE Dominance Function Used to estimate the relative advantage of policy improvement, combined with multi-scenario reward calculation:

[0134] .

[0135] in, For timing difference error, As a discount factor, These are GAE parameters that control the trade-off between bias and variance.

[0136] PPO shearing target Limit the policy update step size to prevent policy collapse due to scenario diversity:

[0137] .

[0138] in, For strategy ratio, This is the cut-off parameter. This mechanism ensures that policy updates occur within the trust domain.

[0139] Expected output of final action probability By weighted fusion of actions from multiple scenarios, deterministic execution instructions are output:

[0140] .

[0141] in, For the scene The derivation implements a mapping from random scenarios to deterministic actions, avoiding over-conservatism in worst-case scenarios.

[0142] Regarding parameter configuration, the Actor network structure is 225-180-160, with the input being the state dimension and the output being the action mean and variance; the Critic network structure is 225-256-128, with the input being the state dimension and the output being the state value; the Actor learning rate is 1×e. −5The learning rate must be less than the Critic learning rate to ensure that the value network converges before the policy network; the Critic learning rate is 1×e −4 Shear parameters The constraint policy ratio is 0.2, and the KL divergence threshold is 0.015, which serves as the trigger condition for the adaptive penalty coefficient. The number of scenes reduced, M, is usually 10-50 to balance the computational load and scene coverage.

[0143] The beneficial effects of the optimization unit are as follows: Initial scenes are generated through GRU probabilistic prediction and Latin hypercube sampling, and representative scenes are retained through synchronous back-sampling, effectively reducing the computational dimensionality of subsequent optimizations; scene probabilities are embedded into the reward function and state space of the Markov decision process, enabling the agent to perceive the uncertainty distribution during training; the shearing and substitution mechanism of proximal policy optimization constrains the policy update step size, achieving optimal scheduling that balances economy and robustness while ensuring monotonic policy improvement, effectively addressing the conservative problem of robust optimization decision-making caused by multi-source uncertainty of the load.

[0144] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0145] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A cloud-edge collaborative intelligent scheduling system, characterized in that, The system includes: The cloud processing unit is used to extract semantic features and fuse graph spatiotemporal structures from industrial system data, and generate prior features for link prediction. The distribution unit is used to perform non-uniform partitioning and encapsulation of the scheduling task containing the prior features based on the total running time model to generate prediction tasks, and to dynamically instantiate and distribute the prediction tasks to the edge gateway. The prediction unit is used to assign a mapping vector to each prediction task to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state, and to solve the shared output layer weights based on the projected state to obtain the multi-task time series prediction results. The optimization unit is used to generate a probabilistic scenario based on the multi-task time-series prediction results, embed the probabilistic scenario into a Markov decision process, and output scheduling action instructions using a policy optimization algorithm.

2. The system according to claim 1, characterized in that, The cloud processing unit includes a semantic extraction module, a graph structure modeling module, and a spatiotemporal evolution analysis module, wherein: The semantic extraction module is used to extract semantic features from the node and edge attributes in the industrial system data through the large language model, and generate semantic embedding vectors. The graph structure modeling module is used to construct a weighted graph structure based on the semantic embedding vector, and to perform spatial aggregation on the weighted graph structure through a graph convolutional network to generate spatial features. The spatiotemporal evolution analysis module is used to perform temporal evolution analysis on the spatial features and generate prior features for link prediction.

3. The system according to claim 2, characterized in that, The graph structure modeling module is specifically used for: Based on the semantic embedding vector and the device connection relationships and operating status sequences in the industrial system data, the weighted graph structure is constructed; The spatial features are generated by spatially aggregating the weighted graph structure using a graph convolutional network. The spatial features are processed through a robust identification mechanism, and the links with high uncertainty are downweighted to filter out noise interference.

4. The system according to claim 3, characterized in that, The spatiotemporal evolution analysis module is specifically used for: The spatial features after robust identification processing are analyzed over time using a time-series modeling algorithm to obtain the prior features for link prediction.

5. The system according to claim 4, characterized in that, The time-series modeling algorithm includes a bidirectional long short-term memory network and a Transformer architecture. The bidirectional long short-term memory network is used to extract local short-range fluctuation features in the spatial features, and the Transformer architecture is used to capture global long-range dependencies and periodic evolution patterns in the spatial features.

6. The system according to claim 1, characterized in that, The distribution unit includes a task partitioning module, specifically used for: Construct a total runtime model that includes startup overhead, communication latency, and computation time; Based on the total runtime model, the scheduling task containing the prior features is dynamically and non-uniformly partitioned and encapsulated to obtain the predicted task, so that the first-executed shards can obtain more resources to mask the startup delay, and the later-executed shards are dynamically adjusted to eliminate tail-end idleness.

7. The system according to claim 6, characterized in that, The distribution unit also includes a resource management module, specifically used for: Based on the non-uniform partitioning results, two-layer dynamic instantiation is performed to establish a distributed computing instance set, and the prediction task is sent to the edge gateway. The first-level instantiation allocates basic container groups based on the divided task fragments, while the second-level instantiation dynamically expands and shrinks based on the length of the real-time task queue.

8. The system according to claim 1, characterized in that, The prediction unit includes a subspace mapping module, specifically used for: For each prediction task, a binary mapping vector with a specific proportion of zero elements is generated, and the binary mapping vector is subjected to a Hadamard product operation with the shared reservoir state to project the shared reservoir state to a low-dimensional task-specific subspace to obtain the projected state.

9. The system according to claim 8, characterized in that, The prediction unit also includes an output solution module, specifically used for: Based on the projection state, ridge regression is used to solve for the shared output layer weights to obtain the time-series prediction results of the prediction task. The multi-task time series prediction results are obtained by summarizing multiple time series prediction results.

10. The system according to claim 9, characterized in that, The optimization unit is specifically used for; Based on the multi-task time-series prediction results, multiple probabilistic scenarios are generated using Latin hypercube sampling, and a representative scenario set is obtained through synchronous back-substitution reduction. The probability of each scenario in the representative scenario set is used as a weight to embed into the reward function of the Markov decision process, and the scenario variables are extended to the state space to construct a Markov decision process with probability weighting. A shearing substitution mechanism with near-end policy optimization is used to constrain the policy update step size, and the policy is updated for the probability-weighted Markov decision process, outputting the scheduling action instruction.

Images