A large model demand forecasting method and system based on deep learning

Abstract

The present invention relates to the field of deep learning technology, and discloses a large-model demand forecasting method and system based on deep learning. The method comprises: screening out a set of trusted nodes that meet a reliability threshold based on the data integrity index and transmission stability index of the Internet of Things device; constructing a decentralized network using the geographic coordinate matrix and communication delay matrix of the trusted node set, and adding noise to the original data of the trusted node set to generate noisy training data; using the noisy training data in the decentralized network to perform federated learning on a first spatiotemporal fusion neural network model to obtain a second spatiotemporal fusion neural network model; inputting real-time data into the second spatiotemporal fusion neural network model to perform feature calculation and demand fluctuation analysis to generate demand forecast results. The present invention can simultaneously capture spatial correlation and temporal dependency features, significantly improving the modeling capability of complex spatiotemporal patterns, and improving the practicality and credibility of the forecast results.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a large-model demand forecasting method and system based on deep learning. Background Art

[0002] Existing demand forecasting systems typically employ a centralized architecture, centralizing data from all IoT devices to a single server for processing. This approach not only presents the risk of single points of failure but is also prone to computational bottlenecks and communication congestion when processing massive amounts of heterogeneous IoT data. Furthermore, traditional methods lack a dynamic assessment mechanism for IoT device reliability, making it impossible to effectively filter out high-quality data sources. This makes forecasting models susceptible to low-quality data and malicious nodes, severely reducing the accuracy and stability of demand forecasts.

[0003] Data privacy protection is an increasingly prominent issue in the IoT environment, and existing demand forecasting methods generally lack effective privacy protection mechanisms. Traditional centralized data processing requires IoT devices to upload raw data directly to a central server. This practice not only violates the privacy protection principle of data localization but also easily leads to sensitive information leakage and privacy violations. While some basic data encryption methods exist in the existing technology, these methods primarily focus on protecting data during transmission and fail to address privacy leaks during data processing and analysis. This is particularly true in the collaborative IoT environment, where personalized privacy protection strategies tailored to the varying sensitivity of different devices are lacking. Summary of the Invention

[0004] The present invention provides a large-model demand forecasting method and system based on deep learning. The present invention can simultaneously capture spatial correlation and temporal dependency features, significantly improving the modeling capability of complex spatiotemporal patterns and increasing the practicality and credibility of forecasting results.

[0005] In a first aspect, the present invention provides a large-model demand forecasting method based on deep learning, the large-model demand forecasting method based on deep learning comprising:

[0006] Filter out a set of trusted nodes that meet the reliability threshold based on the data integrity index and transmission stability index of IoT devices;

[0007] Building a decentralized network using the geographic coordinate matrix and communication delay matrix of the trusted node set, and adding noise to the original data of the trusted node set to generate noisy training data;

[0008] Performing federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model;

[0009] The real-time data in the target IoT environment is input into the second spatiotemporal fusion neural network model for feature calculation and demand fluctuation analysis to generate demand forecast results.

[0010] Optionally, in a first implementation of the first aspect of the present invention, screening out a set of trusted nodes that meet a reliability threshold based on a data integrity index and a transmission stability index of the IoT device includes:

[0011] Real-time monitoring of the data packet integrity rate and transmission delay jitter of IoT devices to obtain data integrity indicators and transmission stability indicators respectively;

[0012] Performing weighted calculation based on the data integrity index and the transmission stability index according to preset weight coefficients to establish a device reliability evaluation matrix containing multiple time windows;

[0013] Based on the device reliability evaluation matrix, the comprehensive trust value of each IoT device is calculated by combining data accuracy, data freshness and data coverage;

[0014] The comprehensive trust value is compared with the reliability threshold, and the IoT devices whose trust values ​​are greater than the reliability threshold are selected as the trusted node set.

[0015] Optionally, in a second implementation of the first aspect of the present invention, constructing a decentralized network using the geographic coordinate matrix and the communication delay matrix of the trusted node set, and adding noise to the original data of the trusted node set to generate noisy training data includes:

[0016] Obtaining the longitude and latitude coordinates of each IoT device in the trusted node set, establishing a geographic coordinate matrix, and constructing a communication delay matrix by collecting network communication delay data between each IoT device;

[0017] Calculating the physical distance value between each IoT device based on the geographic coordinate matrix, and calculating the network connection weight value in combination with the communication delay matrix and the physical distance value;

[0018] Constructing an adjacency matrix based on the network connection weight values, and dynamically adjusting the adjacency matrix according to the time-varying reconnection probability to generate a decentralized network;

[0019] Device-level Laplace noise and aggregation-level Laplace noise are respectively added to the original data of the trusted node set to generate noisy training data.

[0020] Optionally, in a third implementation of the first aspect of the present invention, calculating the physical distance values ​​between the IoT devices based on the geographic coordinate matrix, and calculating the network connection weight values ​​in combination with the communication delay matrix and the physical distance values, includes:

[0021] Extract the latitude and longitude values ​​of each IoT device from the geographic coordinate matrix, and calculate the physical distance between the IoT devices using a spherical distance formula combined with the Earth radius constant;

[0022] The geographical distance influence factor and the communication delay influence factor are set as weight distribution parameters, and the distance standardization parameter and the delay standardization parameter are set as attenuation control variables;

[0023] Calculating a distance weight component through an exponential decay operation based on the physical distance value and the distance normalization parameter, and calculating a delay weight component through an exponential decay operation based on the delay data in the communication delay matrix and the delay normalization parameter;

[0024] The distance weight component is weighted according to the geographic distance impact factor to obtain a first connection weight value, the delay weight component is weighted according to the communication delay impact factor to obtain a second connection weight value, and the first connection weight value and the second connection weight value are added to obtain the network connection weight value between each IoT device.

[0025] Optionally, in a fourth implementation of the first aspect of the present invention, adding device-level Laplace noise and aggregation-level Laplace noise to the original data of the set of trusted nodes to generate noisy training data includes:

[0026] Calculate a device-level privacy budget based on the sensitive data ratio and computing power characteristics of each IoT device in the trusted node set, and calculate an aggregate-level privacy budget based on the training rounds and the maximum privacy budget;

[0027] Generating corresponding Laplace distribution parameters based on the device-level privacy budget value and the global sensitivity parameter, and adding device-level Laplace noise to the original data of each IoT device based on the Laplace distribution parameters to obtain device-level noisy gradient data;

[0028] Aggregating the device-level noisy gradient data by average according to the number of devices to obtain average aggregated data, and generating aggregate-level Laplace distribution parameters based on the aggregate-level privacy budget value and the aggregate sensitivity parameter;

[0029] Based on the aggregation-level Laplace distribution parameters, aggregation-level Laplace noise is added to the data after the average aggregation process to generate noisy training data.

[0030] Optionally, in a fifth implementation of the first aspect of the present invention, performing federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model includes:

[0031] Based on the adjacency matrix of the decentralized network, a graph convolutional network is constructed to process spatial correlation features, and a long short-term memory network is constructed to process temporal dependency features, thereby obtaining a first spatiotemporal fusion neural network model.

[0032] Inputting the noise-added training data into the first spatiotemporal fusion neural network model of each node to perform local training calculation to obtain a local gradient vector;

[0033] Based on the neighbor connection relationship of each node in the decentralized network, the local gradient vector is distributedly aggregated to calculate the gradient difference between neighbor nodes;

[0034] Aggregation weight coefficients are allocated based on the data quality scores and device reliability assessment results of each IoT device. At the same time, malicious nodes whose gradient differences exceed a preset threshold are detected and excluded from the distributed aggregation process, and a second spatiotemporal fusion neural network model is output.

[0035] Optionally, in a sixth implementation of the first aspect of the present invention, performing distributed aggregation on the local gradient vector based on the neighbor connection relationship of each node in the decentralized network and calculating the gradient difference between neighbor nodes includes:

[0036] Determining a neighbor node set of each IoT device node based on a neighbor connection relationship of each node in the decentralized network;

[0037] Each IoT device node broadcasts its own local gradient vector to the set of neighboring nodes, and simultaneously receives gradient vector data from neighboring nodes to construct a comparison matrix between the local gradient vector and the neighboring gradient vector;

[0038] A difference vector between the local gradient vector of each IoT device node and the gradient vector of each neighboring node is calculated based on the comparison matrix, and a gradient difference between the neighboring nodes is generated based on the difference vector.

[0039] Optionally, in a seventh implementation of the first aspect of the present invention, inputting real-time data in the target IoT environment into the second spatiotemporal fusion neural network model to perform feature calculation and demand fluctuation analysis to generate a demand forecast result includes:

[0040] Collect the current sensor data and device status information of each IoT device node in the target IoT environment to build real-time data;

[0041] Inputting the real-time data into the graph convolutional network in the second spatiotemporal fusion neural network model to extract spatial correlation features and generate a spatial feature vector;

[0042] Performing time-dependent feature calculation on the spatial feature vector using the long short-term memory network in the second time-space fusion neural network model to generate a time-series feature vector;

[0043] Performing a matrix operation on the time series feature vector and the IoT device state transition matrix to obtain a fused spatiotemporal correlation feature, performing a nonlinear transformation based on the fused spatiotemporal correlation feature, and outputting a multivariate demand forecast value including a demand quantity forecast, a demand type forecast, and a demand time forecast;

[0044] A demand fluctuation analysis is performed based on the multivariate demand forecast value to generate a demand forecast result.

[0045] Optionally, in an eighth implementation of the first aspect of the present invention, performing demand fluctuation analysis based on the multivariate demand forecast value to generate a demand forecast result includes:

[0046] Building a Markov state transition chain based on the historical operating status data of the IoT devices in the trusted node set, and calculating the probability distribution matrix of each IoT device transitioning from its current state to different target states;

[0047] Decomposing the multivariate demand forecast value according to the demand quantity dimension, demand type dimension and demand time dimension to obtain the demand change rate and demand growth trend coefficient;

[0048] Based on the probability distribution matrix and the demand change rate, the device state evolution path in the future time window is calculated using a Markov prediction algorithm to obtain the impact weight of the device state change on the demand fluctuation;

[0049] Performing weighted correction on the demand growth trend coefficient according to the impact weight, while introducing a demand fluctuation threshold to perform risk assessment and identify the probability interval of abnormal demand fluctuation;

[0050] Based on the probability interval and the confidence parameter, a demand forecast result including a demand forecast value, a forecast confidence interval and a risk level identifier is generated.

[0051] In a second aspect, the present invention provides a large-model demand forecasting system based on deep learning, the large-model demand forecasting system based on deep learning comprising:

[0052] A screening module is used to screen out a set of trusted nodes that meet the reliability threshold based on the data integrity index and transmission stability index of IoT devices;

[0053] A construction module, configured to construct a decentralized network using the geographic coordinate matrix and the communication delay matrix of the trusted node set, and to add noise to the original data of the trusted node set to generate noisy training data;

[0054] a federated learning module, configured to perform federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model;

[0055] The generation module is used to input the real-time data in the target Internet of Things environment into the second spatiotemporal fusion neural network model to perform feature calculation and demand fluctuation analysis to generate demand forecast results.

[0056] The technical solution provided by this invention establishes a dynamic node trust evaluation mechanism, enabling real-time monitoring of IoT device data integrity and transmission stability, dynamic identification and exclusion of unreliable devices, and ensuring the participation of high-quality data sources. A geo-aware decentralized network topology is constructed, integrating device geographic coordinates and communication latency characteristics into the network construction process. This significantly improves communication efficiency and data transmission stability compared to traditional random connection methods. A layered differential privacy protection mechanism overcomes the limitations of traditional single-layer protection by dynamically allocating privacy budgets at both the device and aggregate levels, enabling personalized privacy control. A fully decentralized federated learning architecture completely eliminates reliance on central servers, eliminates single-point failure risks based on a consensus algorithm, and significantly reduces communication overhead. A spatiotemporal fusion neural network, combining graph convolutional networks and long short-term memory networks, can simultaneously capture spatial correlations and temporal dependencies, significantly improving the modeling capabilities of complex spatiotemporal patterns. Intelligent demand fluctuation analysis models device state evolution using a Markov state transition chain, generating comprehensive predictions that include confidence intervals and risk levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0058] Figure 1 A schematic diagram of an embodiment of a large-model demand forecasting method based on deep learning in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of an embodiment of a large model demand forecasting system based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] An embodiment of the present invention provides a large-model demand forecasting method and system based on deep learning. The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0061] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the large model demand forecasting method based on deep learning in an embodiment of the present invention includes:

[0062] Step S101: Filter out a set of trusted nodes that meet a reliability threshold based on the data integrity index and transmission stability index of the IoT device;

[0063] It is understood that the execution subject of the present invention can be a large model demand forecasting system based on deep learning, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0064] Specifically, a real-time monitoring mechanism is constructed for a multi-source dynamic data environment. This mechanism continuously tracks the stability of the communication link and data quality of each IoT device connected to the network. Data link layer communication status information is used to calculate the packet integrity rate of the device at each moment. This statistical result is used to describe the structural integrity and content consistency of the data generated by the device, thereby deriving a data integrity indicator that reflects its data quality. Furthermore, the device's packet transmission delay is frequently detected, and the jitter amplitude and variation trend of this delay are analyzed over multiple sampling periods. This analysis result is used to quantify the stability of the device's network connection as a transmission stability indicator. The data integrity indicator and the transmission stability indicator are weighted according to preset weight coefficients to construct a device reliability assessment matrix covering multiple time windows, in which each item represents the comprehensive communication quality performance of a specific device within a specific time period. A data-dimensional quality evaluation mechanism is introduced, namely, within each time window, the accuracy, freshness, and coverage of the device-collected data are characterized and quantified. Data accuracy is used to assess the degree of deviation between the information collected by the device and the real environment. Data freshness describes the time difference between data generation and use. Data coverage reflects the breadth of the data acquired by the device in terms of spatial area or data category. These three data quality indicators are assigned corresponding weight coefficients and, after weighted combination, form a data quality scoring function. This scoring function is integrated with the results of the device reliability assessment matrix to generate a comprehensive trust value for each device in the current time period. The comprehensive trust values ​​of all devices are compared one by one with the preset reliability threshold. When the trust value of a device is higher than the threshold, it is considered that the device meets the conditions for participating in high-trust distributed learning tasks and is included in the set of trusted nodes.

[0065] Step S102: construct a decentralized network using the geographic coordinate matrix and communication delay matrix of the trusted node set, and add noise to the original data of the trusted node set to generate noisy training data;

[0066] Specifically, a joint modeling of geospatial characteristics and communication performance is performed for all screened trusted IoT devices. The latitude and longitude coordinates of each device are extracted from the trusted node set and populated sequentially according to device number to construct a two-dimensional geographic coordinate matrix, with each row representing the location attribute of a device. Simultaneously, a distributed network diagnostic module performs a large number of ICMP-based ping tests to measure and record the average communication delay of all devices to each other. This generates a symmetric communication delay matrix, which is updated in real time within different time windows and used to characterize the data transmission quality and response characteristics between devices. Based on the Earth's spherical distance formula, spherical trigonometry is performed on each pair of device coordinates in the geographic coordinate matrix to obtain the physical distance between them. This distance is then normalized and merged with the corresponding element in the communication delay matrix. By introducing an exponential decay function, the physical distance and communication delay are weighted and combined with a specific proportional coefficient to calculate a network connection weight matrix. Each element in this matrix represents the strength of the communication connection between two devices. Based on this, a connection weight threshold is set, and the weight matrix is ​​binarized to obtain an adjacency matrix. An element value of 1 in the adjacency matrix indicates a valid connection between the corresponding nodes, while a value of 0 indicates a disconnection. A dynamic reconnection probability model is also introduced. This model incorporates network runtime as a variable into the reconnection probability formula and dynamically adjusts the adjacency matrix based on network evolution. This ensures that the entire network topology exhibits small-world properties and time adaptability, resulting in a decentralized federated learning network architecture with scalable structure, path redundancy, and strong fault tolerance. A two-layer differential privacy protection mechanism is implemented for the raw data in the trusted node set. During the data processing phase, Laplace noise is added locally on the device and at the aggregation node. During the device-level noise generation process, a differential privacy budget is allocated to each device based on its sensitivity level and computational power adjustment factor. The local gradients uploaded by the device are Laplace perturbed based on a global sensitivity parameter to prevent individual data leakage. At the aggregation node level, the noisy gradients uploaded by each device are uniformly re-perturbed based on the current training phase and the global privacy budget decay strategy, ultimately generating noisy training data with a higher level of privacy protection.

[0067] Step S103: performing federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model;

[0068] Specifically, based on a decentralized network adjacency matrix, a graph modeling process is used to model the physical connections and communication structures between nodes. A graph convolutional network (GCN) is then constructed as the backbone module for spatial feature extraction, exploiting the spatial dependencies between devices due to physical proximity or communication stability. To fully capture the temporal evolution and state transition patterns of IoT data, a long short-term memory (LSTM) network is constructed in parallel. This network uses the historical time series collected by each node as input to model the dynamic changes in device states within different time windows, initializing the first spatiotemporal fusion neural network model. The GCN processes the spatial structure between devices, while the LSTM network models the temporal progression of the data. These two components form a coupled structure within the model through a feature fusion mechanism, enabling the joint processing of multi-dimensional, heterogeneous spatiotemporal data. Device- and aggregate-level noise-processed training data is distributed to each trusted node, and each node performs independent model training locally. The local noised samples are used to perform forward propagation, loss function calculation, and gradient backpropagation on the first spatiotemporal fusion neural network model, generating the corresponding local gradient vector. Since the system uses a decentralized federated learning framework, each node will not upload the model parameters to the central server after training is completed. Instead, it will exchange information with its direct neighbor nodes based on the communication path defined by the current network adjacency matrix. During the distributed gradient aggregation process, each node will perform a difference calculation between the local gradient vector calculated by itself and the gradient of its neighboring nodes. The system will gradually perform decentralized weighted aggregation operations based on the local consistency reflected by these gradient differences. At the same time, in order to avoid the interference of low-quality data and abnormal communication on the model training results, data quality scores and equipment reliability assessment results are introduced as aggregation weight distribution factors. The weight of each node is jointly determined by the accuracy, freshness and coverage of its data at the current moment, and is normalized by its average reliability score under multiple time windows. This is used as a quantitative indicator of the influence of each node on its neighbor gradients in the aggregation calculation. A Byzantine fault-tolerance mechanism is introduced during the gradient interaction process, applying a threshold to the gradient differences between all connected nodes. If the gradient deviation of a neighboring node consistently exceeds the system's preset tolerance, the node is deemed to be potentially malicious or experiencing system anomalies and removed from the list of neighbors participating in the current aggregation process, preventing it from misleading the convergence direction of the global model. As distributed training rounds progress and multiple rounds of interaction and aggregation between trusted nodes are completed, parameter synchronization and model consistency are gradually achieved across the entire network, ultimately outputting a trained and optimized second spatiotemporal fusion neural network model.

[0069] Step S104: input the real-time data in the target IoT environment into the second spatiotemporal fusion neural network model to perform feature calculation and demand fluctuation analysis to generate demand forecast results.

[0070] Specifically, the latest operating status information and current sensor data of IoT device nodes distributed at various locations are collected, including dynamic monitoring parameters such as temperature, humidity, voltage, current, flow, and location, as well as static or discrete marker variables such as the device's on / off status, task execution status, and communication status. After collection, the multi-source heterogeneous data is unified into a standardized structural format, forming a real-time data input set with consistent input feature dimensions and synchronized time tags. This real-time data is input into a second spatiotemporal fusion neural network model optimized by federated learning. The graph convolutional network structure in the model is used to model and extract features from the spatial relationships between nodes at the current moment. The graph convolutional layer uses a multi-layer feature propagation mechanism to encode the spatial context of each device node based on the previously established network topology adjacency matrix, as well as the physical distance, communication strength, and functional synergy between nodes, thereby generating a spatial feature vector that reflects the current environmental structural relationships. The spatial feature vector is input into the long short-term memory (LSTM) network of the second spatiotemporal fusion neural network. Using a dynamic control mechanism of memory units and forget gates, the model recursively calculates the temporal evolution trend of each node's spatial features, generating a time-dependent feature vector that captures the changing patterns and potential correlations between device operating states in consecutive time segments. This time-dependent feature vector represents the current node's behavioral trajectory over a historical period and possesses excellent dynamic predictive capabilities. To enhance the model's logical reasoning capabilities during state prediction, a state transition matrix is ​​introduced and matrix multiplied with the time-dependent feature vector to construct a spatiotemporal joint feature representation that incorporates the dynamic evolution path of the device's operating state. Furthermore, the fused spatiotemporal features are subjected to feature mapping and dimensionality reduction through a multi-layer nonlinear transformation structure, resulting in a structured multivariate demand forecast. This forecast contains three dimensions: the demand quantity forecast for a particular device or region in the future time period, the demand type forecast label, and the predicted time point at which the demand will occur. Demand fluctuation analysis is performed based on the multivariate forecast outputs of all devices. By identifying amplitude changes, frequency changes, and sudden increase patterns in the forecast data, potential load concentration trends, capacity boundary pressures, and abnormal behavior prompts are extracted to form the final demand forecast results.

[0071] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0072] Real-time monitoring of the data packet integrity rate and transmission delay jitter of IoT devices to obtain data integrity indicators and transmission stability indicators respectively;

[0073] Based on the data integrity index and transmission stability index, a weighted calculation is performed according to the preset weight coefficient to establish an equipment reliability evaluation matrix containing multiple time windows;

[0074] Based on the device reliability evaluation matrix, the comprehensive trust value of each IoT device is calculated by combining data accuracy, data freshness and data coverage;

[0075] The comprehensive trust value is compared with the reliability threshold, and IoT devices with trust values ​​greater than the reliability threshold are selected as the trusted node set.

[0076] Specifically, the system deploys a data monitoring agent module on each IoT device node. This module, running in the device firmware or access gateway, performs real-time analysis of downlink data packet sequence numbers, checksum fields, and packet loss records. It calculates the integrity rate of the device's uploaded data within the current time window. This integrity rate is defined as the ratio of the number of packets successfully received and passing content verification per unit time to the theoretical total number of packets transmitted. This rate serves as the node's data integrity indicator at the current moment. The system frequently monitors the real-time fluctuations in uplink and downlink communication delays on each device, calculates the standard deviation of the arrival time intervals of multiple consecutive packets, and uses a sliding window algorithm to filter out abnormal spikes. This results in a transmission delay jitter signature for each device within the current period. This signature constitutes the device's transmission stability indicator, which measures the stability of network link quality on a microscopic timescale. The data integrity and transmission stability indicators are then integrated according to a pre-defined weighting rule. A multidimensional device reliability assessment matrix is ​​constructed based on the device's performance within different time windows. This matrix is ​​structured with device number as the row dimension and time window number as the column dimension. Each element in the matrix is ​​calculated using a set of weighted combination functions. Based on the device reliability assessment matrix, the system calculates the comprehensive trustworthiness of each IoT device by combining data accuracy, data freshness, and data coverage. Data accuracy measures the deviation between actual measurements and theoretical or reference values ​​and is applicable to monitoring physical quantities within a standard range, such as temperature and humidity, voltage and current, or GPS location. Data freshness is determined by the difference between the data generation timestamp and the system reception timestamp, reflecting the latency during data transmission and processing within the network. Data coverage assesses the representativeness and completeness of the data collected by the device within the spatial, dimensional, or categorical distribution range to assess its representativeness and completeness within the overall network or mission objectives. The system assigns weights to each of these three metrics and combines them into a unified data quality score using a weighted aggregation function. This score is then combined with the time window average of each device in the device reliability assessment matrix to determine the current comprehensive trustworthiness of each device. After calculating the comprehensive trustworthiness of all nodes, the system invokes the trust screening logic module to compare each device's trustworthiness against a pre-set reliability threshold. This threshold is set based on the tolerance requirements for data accuracy, network reliability, and resource budget in actual application scenarios, with a certain degree of flexibility in adjusting upper and lower limits. When a device's overall trust value exceeds this threshold, the system deems the node a valid participant in the current task cycle and marks it as a trusted node. Devices below this threshold are temporarily excluded due to potential instability or low-value data to prevent them from interfering with the distributed model training process. Through this entire process, the system dynamically selects a set of IoT devices that meet the trustworthiness constraints.

[0077] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0078] Obtain the longitude and latitude coordinates of each IoT device in the trusted node set, establish a geographic coordinate matrix, and construct a communication delay matrix by collecting network communication delay data between each IoT device;

[0079] Calculate the physical distance between IoT devices based on the geographic coordinate matrix, and calculate the network connection weight value by combining the communication delay matrix and the physical distance value;

[0080] Construct an adjacency matrix based on the network connection weight values, and dynamically adjust the adjacency matrix according to the time-varying reconnection probability to generate a decentralized network;

[0081] Device-level Laplace noise and aggregate-level Laplace noise are added to the original data of the trusted node set to generate noisy training data.

[0082] Specifically, the system extracts geolocation parameters from highly trusted IoT devices. The longitude and latitude values ​​of each device are sampled, denoised, and formatted using a standardized encoding method to form a location information sequence. The system then arranges this sequence by device number, constructing a two-dimensional geographic coordinate matrix. Each row of the matrix corresponds to the geographic coordinate point of a trusted device, specifically a two-dimensional vector consisting of the device's latitude and longitude values. Simultaneously, the system activates the device communication link status detection module. By performing large-scale bidirectional ping tests and UDP packet exchange tests between device pairs, the system collects round-trip communication delays within different time windows. The observed average values ​​are then populated into the communication delay matrix, forming a symmetrical n×n structure, where n is the number of trusted devices. Each element in the matrix reflects the average network delay between two devices. Based on the geographic coordinate matrix and the communication delay matrix, the system performs spatial geometry calculations and time delay correlation fusion operations. Using Earth spherical trigonometry formulas, the system calculates the physical geographic distances between all trusted devices. Specifically, the spherical cosine formula is used to convert longitude and latitude coordinates into spherical angles. These angles are then multiplied by the Earth's radius constant to obtain the geographic distance between each pair of devices, thereby constructing a physical distance matrix. The system normalizes the physical distance matrix and the aforementioned communication delay matrix to eliminate the effects of scale differences. It also introduces an exponential weight decay function, embedding both into a network connection weight calculation model. This model comprehensively considers spatial density and communication quality stability, assigning higher connection weights to device pairs with closer physical proximity and lower communication latency, while assigning exponentially decayed weak connection weights to connections with severe spatial separation or high latency. This results in a network connection weight matrix that reflects the closeness of connections between device pairs. Based on this connection weight matrix, the system employs a weight threshold determination mechanism to construct an initial adjacency matrix. Device pairs with weights above the threshold are marked as 1, indicating a logical connection, and connections with weights below the threshold are marked as 0, indicating no connection. This adjacency matrix directly constitutes the edge set representation of the first version of the decentralized network topology. To overcome topological instabilities caused by node state fluctuations, temporary changes in communication links, and physical layout adjustments in IoT environments, the system introduces a time-driven reconnection probability function to dynamically reconstruct and adaptively adjust the adjacency matrix. This reconnection probability function uses time t as a variable and is determined by setting the maximum reconnection probability ρ and the evolution time constant. An exponential growth function is constructed to ensure structural stability in the early stages of network operation, while gradually increasing the degree of freedom for network reconnection in later stages. At each clock cycle, the system calculates the probability of updating the current connection based on the persistence of the current connection, changes in trust scores, and latency growth trends. The system dynamically adjusts the connection state in the adjacency matrix under probability threshold control, thereby forming a decentralized network graph structure that is self-evolving, self-regulating, and exhibits small-world properties. Differential privacy processing is performed on local raw data at each trusted node to generate privacy-preserving noisy training data. The system allocates a unique device-level privacy budget to each device based on the sensitivity level of the data uploaded by the device in historical tasks and the computing resource performance. Using a Laplace distributed random number generator with the global gradient sensitivity as the scale parameter, the system calculates a device-level Laplace noise vector and superimposes it on the locally calculated gradient value to form the perturbed local model update data. Before the noisy gradients uploaded by the device reach neighboring nodes for model aggregation, the system performs a unified aggregation-level noise injection on multiple gradients within the neighborhood. In this stage, the system uses the aggregate privacy budget to calculate the aggregation-level budget value based on the proportional relationship between the current global training round and the total training cycle, and once again generates the aggregation-level Laplace noise vector, which is superimposed on the local aggregation result to form the final noisy training dataset.

[0083] In a specific embodiment, the process of calculating the physical distance values ​​between the IoT devices based on the geographic coordinate matrix and calculating the network connection weight values ​​in combination with the communication delay matrix and the physical distance values ​​may specifically include the following steps:

[0084] Extract the latitude and longitude values ​​of each IoT device from the geographic coordinate matrix, and calculate the physical distance between the IoT devices using the spherical distance formula combined with the Earth's radius constant;

[0085] The geographical distance influence factor and the communication delay influence factor are set as weight distribution parameters, and the distance standardization parameter and the delay standardization parameter are set as attenuation control variables;

[0086] Calculate the distance weight component through exponential decay operation based on the physical distance value and the distance normalization parameter, and calculate the delay weight component through exponential decay operation based on the delay data in the communication delay matrix and the delay normalization parameter;

[0087] The distance weight component is weighted according to the geographical distance influence factor to obtain a first connection weight value, the delay weight component is weighted according to the communication delay influence factor to obtain a second connection weight value, and the first connection weight value and the second connection weight value are added together to obtain the network connection weight value between each IoT device.

[0088] Specifically, the geographic location information of all IoT devices is extracted from the trusted node geographic coordinate matrix, where the location of each node is represented by a two-dimensional vector consisting of latitude and longitude values. For any two devices, the system takes their corresponding latitude and longitude, and then derives the great-circle distance between them using the spherical cosine theorem. This spherical distance calculation process uses the Earth's radius as a constant and combines trigonometric functions to form a complete path, thereby obtaining the actual geographic distance between the two devices in physical space. The average network delay value between any two nodes is extracted from the communication delay matrix as the basic data for delay evaluation. The system enters the connection weight value modeling phase, in which two core weight adjustment factors are set: the geographic distance influence factor and the communication delay influence factor. These two coefficients reflect the respective importance of spatial factors and network performance in connection modeling. At the same time, the distance normalization parameter and the delay normalization parameter are introduced as variables to control their respective influence range and attenuation speed, respectively, to support flexible adjustment of the nonlinear attenuation function. When calculating connection weights, an exponential decay calculation is performed on the physical distance values ​​of all node pairs. Based on the distance normalization parameter, the system uses the physical distance as input and, through a decaying transformation, generates a geographic distance weight component between each pair of devices. This component approaches its maximum value when the nodes are close together and rapidly approaches its minimum value when the distance between them is much greater than the set parameter. This results in strong connections between spatially adjacent nodes and weak connections or even insignificance between spatially separated nodes. Similarly, the system applies an exponential decay operation to the delay data in the communication delay matrix, calculating the corresponding decay result based on the delay normalization parameter. This result reflects the regulatory effect of network path stability on connection strength: low-latency connections receive high weights, while high-latency connections receive rapidly decaying low weights, reflecting the system's real-time response to communication quality. The two weight components are weighted according to their corresponding influence factors: the geographic distance weight component is multiplied by the geographic distance influence factor to obtain the first connection weight, and the communication delay weight component is multiplied by the communication delay influence factor to obtain the second connection weight. For example, in industrial control systems with higher requirements for communication reliability, the communication delay influence factor is higher than the geographic distance influence factor, while in low-power, long-distance monitoring networks, the geographic distance influence factor dominates. After the weighting process is completed, the first connection weight value and the second connection weight value are numerically added, that is, the two connection strength sources of each pair of devices are accumulated to obtain a network connection weight value reflecting the comprehensive connection capability between the device pair.

[0089] In a specific embodiment, the step of adding device-level Laplace noise and aggregate-level Laplace noise to the original data of the trusted node set, and the process of generating the noisy training data may specifically include the following steps:

[0090] The device-level privacy budget is calculated based on the sensitive data ratio and computing power characteristics of each IoT device in the trusted node set, and the aggregate-level privacy budget is calculated based on the training rounds and the maximum privacy budget.

[0091] Based on the device-level privacy budget value and the global sensitivity parameter, the corresponding Laplace distribution parameters are generated. Based on the Laplace distribution parameters, device-level Laplace noise is added to the original data of each IoT device to obtain device-level noisy gradient data.

[0092] Aggregate the device-level noisy gradient data by the number of devices to obtain the average aggregated data, and generate the aggregate-level Laplace distribution parameters based on the aggregate-level privacy budget value and the aggregate sensitivity parameter.

[0093] Based on the aggregation-level Laplace distribution parameters, aggregation-level Laplace noise is added to the averaged aggregation data to generate noisy training data.

[0094] Specifically, the system dynamically analyzes the local data characteristics of all IoT devices in the trusted node set and extracts two key metrics: the proportion of sensitive data involved in each device's current task cycle and the device's computing power. The sensitive data proportion is calculated as the ratio of the amount of data uploaded by the device labeled as sensitive by the system to its total data volume, while the computing power characteristic is derived from a comprehensive assessment of the device's computing power, storage, bandwidth, and other parameters. These two metrics directly determine the privacy budget strength and noise tolerance assigned to the device by the system. In specific implementation, the system calculates a device-level privacy budget for each device using a composite mapping rule based on a conditioning function set for the sensitive data proportion and a compensation factor corresponding to the computing power. This value, a positive number, represents the maximum perturbation strength that the device can tolerate in its local gradient calculation results and also affects the magnitude of its Laplace noise injection. Simultaneously, the system initiates an aggregate-level privacy budget control strategy. Based on the current model training epoch and the system-set maximum privacy budget cap, an aggregate budget decay model is constructed using a budget decay factor to calculate the aggregate-level privacy budget applicable to the global model aggregation operation at the current moment. This budget value is round-dependent: maintaining a high budget in the early stages of training to enhance model learning and gradually reducing it later to improve model stability and generalization. The device-level privacy budget is combined with the preset global sensitivity parameter to construct a Laplace distribution control scale. This generates independent Laplace distribution parameters for each device. These parameters define the width of the probability density function of the noise perturbation, indirectly controlling the impact of the perturbation on model update accuracy. Based on these distribution parameters, the system randomly samples from the corresponding Laplace probability density function to generate a symmetric perturbation vector. This perturbation vector is then superimposed on the gradient vector calculated by the device during local training to produce noisy gradient data protected against device-level differential perturbations. The system collects the device-level noisy gradient data uploaded by all devices and performs the first stage of weighted aggregation. This operation averages the noisy gradient data of each device based on the number of devices, producing an aggregated intermediate gradient result that represents the direction of the model update for this round. The system invokes the aggregate-level privacy budget control module to generate global Laplace distribution parameters based on the aggregate-level privacy budget value for the current round and the system-defined aggregation sensitivity parameters. The system then applies the aggregate-level Laplace noise injection mechanism to the intermediate gradient data after average aggregation. The system resamples the perturbation vector based on the generated Laplace distribution parameters and superimposes this perturbation on the intermediate aggregate data, completing the final level of differential privacy protection and generating noisy training data.

[0095] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0096] Based on the adjacency matrix of the decentralized network, a graph convolutional network is constructed to process spatial correlation features, and a long short-term memory network is constructed to process temporal dependency features, resulting in the first spatiotemporal fusion neural network model.

[0097] Input the noise-added training data into the first spatiotemporal fusion neural network model of each node for local training calculation to obtain a local gradient vector;

[0098] Based on the neighbor connection relationship of each node in the decentralized network, the local gradient vector is distributedly aggregated and the gradient difference between neighboring nodes is calculated;

[0099] Aggregation weight coefficients are allocated based on the data quality scores and device reliability assessment results of each IoT device. At the same time, malicious nodes whose gradient differences exceed the preset threshold are detected and excluded from the distributed aggregation process, and the second spatiotemporal fusion neural network model is output.

[0100] Specifically, an adjacency matrix is ​​extracted based on a decentralized network structure. This adjacency matrix reflects the effective connectivity between devices in the current set of trusted nodes and indicates the set of neighboring devices to which each device is connected. Based on this adjacency matrix, the system designs and constructs a graph convolutional network (GCN) to process the spatial correlation features between IoT devices. By performing weight propagation and structural normalization on adjacency relationships, the GCN models cross-node spatial dependencies. This allows each node to not only learn its own feature representation but also integrate the state information of its immediate neighbors, capturing structural dependencies at the physical or functional level. Furthermore, to adapt to the dynamic nature of IoT data, which is highly dependent on time series features, a long short-term memory (LSTM) network is constructed in parallel. This inputs historical observations of each node into a recurrent neural network structure. Through a joint control mechanism of input, forget, and output gates, the GCN learns the evolutionary path between temporal states and outputs a dynamic feature vector with contextual memory capabilities. The spatial structural features output by the GCN are integrated with the temporal dependency features extracted by the LSTM network through a feature fusion module to form the first spatiotemporal fusion neural network model. After model construction is complete, the system distributes noisy training data to each trusted node. Each node then locally performs forward inference and gradient backpropagation of the first spatiotemporal fusion neural network model. Each node independently trains network parameters using local data and calculates the corresponding gradient vectors, generating local gradient outputs. These local gradients reflect each node's parameter optimization recommendations for the global model based on its private data. After local training is complete, the system organizes each node to perform distributed gradient aggregation based on the connectivity structure of the decentralized network, organized around the connections defined by the adjacency matrix. Each node subtracts its own gradient from the gradients of all its neighbors to assess the consistency of update directions among neighboring nodes in the model parameter space. If the difference between a neighbor's gradient and the local gradient is significantly greater than the average gradient deviation between other connected nodes in the network, and no directional regression occurs over multiple training rounds, the system determines that the node has engaged in potential data contamination, gradient tampering, or communication deception. The node is marked as malicious and removed from the current round of distributed gradient aggregation to ensure that the model convergence direction is not misguided. Through an aggregation weighting mechanism, each node's score in the data quality scoring system and the device's reliability assessment results over a historical period are used to assign a weight coefficient applicable to the current aggregation task. The data quality score is calculated based on the accuracy, freshness, and coverage of the data, reflecting the actual contribution of the device's uploaded data. The reliability assessment focuses on the device's communication stability, normal response rate within the operating cycle, and behavioral stability in collaborative tasks. The scores of these two dimensions are normalized and combined to form the device's weight factor, which controls the proportion of the device's uploaded gradients in the final aggregation result.When the gradient vectors and their corresponding weight factors of all participating nodes are determined, the system performs a weighted aggregation operation according to the decentralized network structure, generating a new model parameter update result. After the distributed aggregation is completed and the parameters of the local models of all nodes are synchronized, the system obtains the second spatiotemporal fusion neural network model.

[0101] In a specific embodiment, the step of performing distributed aggregation of local gradient vectors based on the neighbor connection relationship of each node in the decentralized network and calculating the gradient difference between neighbor nodes may specifically include the following steps:

[0102] Determine the neighbor node set of each IoT device node based on the neighbor connection relationship of each node in the decentralized network;

[0103] Each IoT device node broadcasts its own local gradient vector to the set of neighboring nodes, and at the same time receives gradient vector data from neighboring nodes to construct a comparison matrix between the local gradient vector and the neighboring gradient vector;

[0104] The difference vector between the local gradient vector of each IoT device node and the gradient vector of each neighboring node is calculated based on the comparison matrix, and the gradient difference between the neighboring nodes is generated based on the difference vector.

[0105] Specifically, based on the adjacency matrix structure of a decentralized network, the connection relationships of all trusted device nodes are analyzed one by one, and the set of neighbor nodes for each node in the current network topology is determined accordingly. The system traverses each row in the adjacency matrix, where an element with a value of 1 indicates that the node has a valid connection with the corresponding index node. The system then summarizes the numbers of these connected nodes into a neighbor node index list and establishes an independent neighbor node set for each device node. Each IoT device node broadcasts its local gradient vector to the neighbor node set. That is, after completing the current round of local model training, each IoT device node calculates its own local gradient vector and broadcasts it to all nodes listed in its neighbor node set via a secure communication channel. This broadcast behavior does not require transit through a central server, but instead relies on a point-to-point direct communication mechanism between neighbors to ensure the structural independence of distributed training and minimize communication latency. Simultaneously, each node synchronously receives local gradient vectors from all of its neighboring nodes. These received gradient vectors are labeled according to the neighbor node number and temporarily stored in a local cache. The system then inputs these vectors, along with its own local gradient vector, into a comparison structure, thereby constructing a complete data structure containing the "local gradient - neighbor gradient" relationship. This structure is known as the comparison matrix between local gradient vectors and neighbor gradient vectors. This comparison matrix is ​​a two-dimensional structure oriented in vector space, where the rows represent the node's local target dimension, and the columns represent the gradient values ​​from each neighbor node along the corresponding dimension. A separate column on the left side of the matrix records the local node's original gradient value along the same dimension. This structured organization allows the system to efficiently perform local gradient difference calculations without compromising node privacy. Based on this comparison matrix, the system performs element-by-element vector difference calculations, calculating the difference between the gradient vector of each neighbor node and the local node's own gradient vector along the corresponding dimension, generating a set of difference vectors equal to the number of neighbors. Each difference vector represents the numerical difference between the local model update direction of a specific neighboring node and the current node in this round of training. These difference data are reflected in the degree of deviation in the model training direction in terms of quantity, and structurally reveal the convergence consistency of the local model. The system stores each difference vector and organizes it into a gradient difference matrix between neighboring nodes. For example, when the gradient difference of a neighboring node continues to maintain a high deviation in multiple rounds, and its change trend does not conform to the direction of decreasing global model loss, the system will include it in the abnormal node identification list based on the abnormal behavior record of the node in the comparison matrix, mark its status as suspicious or low credibility, and lower its participation weight in subsequent training rounds, or temporarily remove its participation rights to enhance the robustness of the overall model.

[0106] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0107] Collect the current sensor data and device status information of each IoT device node in the target IoT environment to build real-time data;

[0108] The real-time data is input into the graph convolutional network in the second spatiotemporal fusion neural network model to extract spatial correlation features and generate spatial feature vectors;

[0109] Through the long short-term memory network in the second spatiotemporal fusion neural network model, the time-dependent characteristics of the spatial feature vector are calculated to generate a time-series feature vector;

[0110] Perform matrix operations on the time series feature vector and the IoT device state transition matrix to obtain fused spatiotemporal correlation features. Based on the fused spatiotemporal correlation features, a nonlinear transformation is performed to output a multivariate demand forecast value that includes demand quantity forecast, demand type forecast, and demand time forecast.

[0111] Conduct demand fluctuation analysis based on multivariate demand forecast values ​​and generate demand forecast results.

[0112] Specifically, various device nodes deployed in the target IoT environment serve as collection terminals, continuously acquiring sensor data and device operating status information from each node at preset intervals. Sensor data includes continuous values ​​of various environmental and physical variables, such as temperature, humidity, voltage, current, vibration, noise, light, and gas concentration. Device status information includes discrete and enumerated parameters, such as whether the device is powered on, operating mode number, operating load level, remaining battery life, and communication connection status. Together, these parameters constitute important input features reflecting changes in individual devices and their surrounding environments. After data collection is complete, all data undergoes processing steps, including timestamp unification, missing value repair, and dimensionality normalization, to form a real-time data matrix. This real-time data matrix is ​​then fed into a second spatiotemporal fusion neural network model, which has completed decentralized federated training and achieved stable convergence, and then into a graph convolutional network module to extract spatial structural correlation features between IoT devices. At this stage, the system uses a graph convolutional network to perform weighted propagation operations on the information between each node and its neighbors, based on the previously constructed adjacency matrix and the physical topology or functional synergy between devices. This encodes the node's spatial contextual information in a high-dimensional feature space. Through the feature aggregation mechanism of the graph convolutional layer, each node no longer relies solely on its own sensor data and state, but instead is able to perceive the state trends of its neighboring nodes. This is reflected in the feature representation as a high-dimensional vector embedded in the spatial structure. These spatial feature vectors capture the information environment and collaborative state of the node within its local network structure. After completing the spatial feature extraction, the system passes the spatial feature vector output by the graph convolutional network as a time series input to the long short-term memory (LSTM) module in the second spatiotemporal fusion neural network. In this module, the system takes the spatial feature stream of several consecutive time periods as input. Using memory units and a gating mechanism, it models the long-term dependencies and short-term fluctuations in the time series, gradually capturing the state evolution path and trend characteristics of the device over the past period. The LSTM network, through its recurrent connection structure, automatically selects which historical information to retain and which transient perturbations to forget, thereby generating a time series feature vector that aggregates the temporal context. This time series feature vector is then matrix multiplied with the state transition matrix derived from the historical modeling of the device's operating state. The state transition matrix records the probability distribution of a device's transitions between different operating states. By combining it with time series features, the current state dynamics are integrated with known state evolution patterns, constructing a spatiotemporal feature representation that fuses spatial correlation, temporal memory, and state evolution probability. The fused spatiotemporal correlation features are then input into a nonlinear mapping structure, such as a multilayer perceptron or a deep fully connected layer with an activation function. The nonlinear function enables high-order feature interaction and pattern decoupling, ultimately outputting a prediction result. This nonlinear transformation outputs a structured multivariate representation, encompassing demand quantity forecasts, demand type forecasts, and demand time forecasts.The demand quantity forecast is a continuous value, representing the intensity of resource or task requests at a specific point in time. The demand type forecast is presented as a categorical vector, indicating the type of demand occurring at that node, such as energy requests, maintenance alerts, data uploads, and task collaboration. The demand time forecast is presented as a timestamp or time window indicator, indicating the expected time range or peak time of these demand events. The forecast results from these three dimensions together constitute the multivariate demand forecast output. Based on the multivariate forecast results, the system analyzes demand fluctuations. Combining historical fluctuation patterns with the current forecast amplitude, it analyzes the amplification of short-term trends and assesses the frequency and amplitude of medium- and long-term trends to identify periodic overloads, potential concentrated outbreaks, or window periods. Based on this fluctuation analysis, a structured demand forecast report or dynamic control strategy is generated, which is fed back into actual control processes such as equipment scheduling, energy consumption allocation, load balancing, and task triggering, thereby implementing an intelligent forecasting and closed-loop control mechanism based on deep learning and graph spatiotemporal modeling.

[0113] In a specific embodiment, the execution step performs demand fluctuation analysis based on the multivariate demand forecast value, and the process of generating the demand forecast result may specifically include the following steps:

[0114] Based on the historical operating status data of IoT devices in the trusted node set, a Markov state transition chain is constructed to calculate the probability distribution matrix of each IoT device transitioning from its current state to different target states.

[0115] Decompose the multivariate demand forecast value according to the demand quantity dimension, demand type dimension and demand time dimension to obtain the demand change rate and demand growth trend coefficient;

[0116] Based on the probability distribution matrix and the demand change rate, the Markov prediction algorithm is used to calculate the equipment state evolution path in the future time window, and the impact weight of equipment state change on demand fluctuation is obtained;

[0117] The demand growth trend coefficient is weighted and modified according to the impact weight. At the same time, a demand fluctuation threshold is introduced for risk assessment to identify the probability range of abnormal demand fluctuations.

[0118] Based on the probability interval and confidence parameters, a demand forecast result is generated, which includes the demand forecast value, forecast confidence interval and risk level identification.

[0119] Specifically, the system systematically collects and structures the state data generated by each trusted device node during long-term operation. This state data includes sequence information with discrete state identification characteristics, such as the device's operating mode, switch status, operating load, power level, and environmental response status within each time period. The system sorts this state information sequentially based on timestamps and labels the states within each time period to construct a state sequence sample library. The system uses Markov chain modeling to statistically analyze this state sequence, identifying transitions between states and calculating the frequency of transitions from any state to each possible target state. A state transition probability matrix is ​​obtained through normalization. Each row of this matrix represents the conditional probability of the device entering all possible target states at the next time step under a given initial state. The matrix as a whole forms a stable Markov state transition chain. The multivariate demand forecast value is dimensionally decomposed according to its output structure, extracting three key pieces of information: demand quantity forecast value, demand type forecast label, and demand time forecast window. The demand quantity dimension reflects the system's resource consumption or task call intensity per unit time. The system calculates the demand change rate through the trend of demand value changes at consecutive moments to evaluate whether the current demand is in an upward, downward, or oscillating state. The demand type dimension provides information on changes in the functional direction. Through the changes in the frequency and proportion of different types of demand over time, the system identifies whether there is a clustering trend of requests for specific functions. The demand time dimension is used to calculate the forecast concentration and response advance, and construct the time distribution characteristics of the demand. After combining the above three dimensions, the system calculates the demand growth trend coefficient, which quantifies the overall rising speed and growth slope of demand in the future forecast stage. The system jointly models the Markov state transition probability matrix with the above demand change rate parameters, executes the Markov prediction algorithm to derive the state evolution path of each device in the future time window, and quantifies the contribution weight of different state transition paths to the overall demand fluctuation. At each state transition step in the Markov chain, the system maps the predicted state sequence to the corresponding demand growth trend, numerically modeling the impact of state changes on resource utilization, service triggering, or data upload frequency. The system also calculates the magnitude and direction of the impact of each state transition path on system demand changes under current environmental parameters, thereby generating a weight matrix for the impact of device state changes on overall demand fluctuations. This matrix describes the extent to which the behavior of "devices transitioning from state A to state B" will drive demand up, down, or maintain, thereby expanding the demand forecasting process from simple time series model reasoning to dynamic path backtracking and forward-looking deduction driven by state behavior.Based on the aforementioned influence weight matrix, the system modifies the initial demand growth trend coefficient. Specifically, based on the influence weights assigned to each state path, it applies adjustment factors to the trend parameters for different time periods during the forecast phase. This ensures that the final trend value reflects not only the historical slope of change but also the dynamic regulation effect driven by the state evolution logic. Furthermore, after this modification, the system introduces a demand fluctuation threshold for risk identification and anomaly assessment. This fluctuation threshold is set by the system based on multiple factors, including device type, service characteristics, historical tolerance range, and task response time. The system identifies potential risk areas within intervals where the predicted demand change rate exceeds this threshold and inserts an abnormal fluctuation indicator into the corresponding forecast data. Based on the probability distribution range and corresponding gradient of each path in the Markov transition chain, combined with the confidence output parameter in the forecast model, the system calculates the confidence interval of the forecast output result, constructing a demand forecast result structure with a probability distribution hierarchy. Based on the probability interval and confidence parameter, a demand forecast result is generated, which includes the demand forecast value, the forecast confidence interval, and a risk level indicator.

[0120] The above describes the large model demand forecasting method based on deep learning in the embodiment of the present invention. The following describes the large model demand forecasting system based on deep learning in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a large model demand forecasting system based on deep learning includes:

[0121] A screening module 201 is configured to screen out a set of trusted nodes that meet a reliability threshold based on the data integrity index and transmission stability index of the IoT device;

[0122] A construction module 202 is configured to construct a decentralized network using the geographic coordinate matrix and the communication delay matrix of the trusted node set, and to add noise to the original data of the trusted node set to generate noisy training data;

[0123] A federated learning module 203 is configured to perform federated learning on the first spatiotemporal fusion neural network model using the noisy training data in a decentralized network to obtain a second spatiotemporal fusion neural network model;

[0124] The generation module 204 is used to input the real-time data in the target Internet of Things environment into the second spatiotemporal fusion neural network model to perform feature calculation and demand fluctuation analysis to generate demand forecast results.

[0125] Through the collaborative efforts of these components, a multi-dimensional device reliability assessment system is established by real-time monitoring of IoT device data integrity and transmission stability indicators. Compared to traditional static node selection methods, this system can dynamically identify and exclude unreliable devices, ensuring that participating nodes have high-quality data sources. A time-window-based device reliability assessment matrix captures the time-varying nature of device performance, avoids the limitations of fixed weight assignment in traditional methods, and improves the adaptability and accuracy of node selection. Incorporating the geographic coordinate matrix and communication delay matrix of IoT devices into the network topology construction process significantly improves network communication efficiency and data transmission stability compared to traditional decentralized networks with random connections. A connection weight calculation mechanism based on the spherical distance formula and a decay function constructs a network topology with small-world properties, ensuring network connectivity while reducing communication overhead and optimizing the performance of decentralized collaboration. This system overcomes the limitations of traditional single-layer privacy protection by implementing a two-layer dynamic privacy budget allocation strategy at both the device and aggregate levels, enabling personalized privacy protection based on the sensitive data ratio and computing power characteristics of different IoT devices, achieving more refined privacy control. An aggregation-level privacy budget calculation method based on an exponential decay function dynamically adjusts the strength of privacy protection during training, ensuring data privacy while maximizing model performance, thus balancing privacy protection with model accuracy. This approach completely eliminates reliance on a central server and achieves truly decentralized model aggregation based on a consensus algorithm, eliminating single points of failure and improving system reliability and fault tolerance. Compared to traditional centralized aggregation methods, this distributed aggregation mechanism, which utilizes neighbor connectivity and iterative averaging of gradient differences, significantly reduces communication overhead and computational complexity, making it suitable for large-scale IoT deployments. Combining a graph convolutional network with a long short-term memory network, this approach simultaneously captures both spatial correlation and temporal dependency in IoT data, significantly improving its ability to model complex spatiotemporal patterns compared to traditional single-dimensional prediction models. The introduction of a device state transition matrix and a demand fluctuation prediction function enables modeling the impact of dynamic device state changes on demand in IoT environments, enabling more accurate multi-dimensional demand forecasting. The device state evolution modeling based on the Markov state transition chain can predict the device state change path within the future time window. Through the weighted correction calculation and risk assessment mechanism of the influencing weights, it can generate a comprehensive prediction result including confidence intervals and risk levels, thereby improving the practicality and credibility of the prediction results.

[0126] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0127] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a large-scale demand forecasting device based on deep learning (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program code.

[0128] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large model demand forecasting method based on deep learning, characterized in that: include: Filter out a set of trusted nodes that meet the reliability threshold based on the data integrity index and transmission stability index of IoT devices; Building a decentralized network using the geographic coordinate matrix and communication delay matrix of the trusted node set, and adding noise to the original data of the trusted node set to generate noisy training data; Performing federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model; Collect current sensor data and device status information of each IoT device node in the target IoT environment to construct real-time data; input the real-time data into the graph convolutional network in the second spatiotemporal fusion neural network model to extract spatial correlation features and generate a spatial feature vector; use the long short-term memory network in the second spatiotemporal fusion neural network model to calculate the time-series dependency features of the spatial feature vector to generate a time-series feature vector; perform matrix operations on the time-series feature vector and the IoT device state transition matrix to obtain fused spatiotemporal correlation features, and perform nonlinear transformation based on the fused spatiotemporal correlation features to output a multivariate demand forecast value including demand quantity forecast, demand type forecast, and demand time forecast; A Markov state transition chain is constructed based on the historical operating status data of the IoT devices in the trusted node set, and a probability distribution matrix of each IoT device transitioning from a current state to different target states is calculated; the multivariate demand forecast value is decomposed according to the demand quantity dimension, demand type dimension, and demand time dimension to obtain the demand change rate and the demand growth trend coefficient; based on the probability distribution matrix and the demand change rate, a Markov prediction algorithm is used to calculate the device state evolution path in the future time window to obtain the impact weight of the device state change on the demand fluctuation; the demand growth trend coefficient is weighted and corrected according to the impact weight, and a demand fluctuation threshold is introduced for risk assessment to identify the probability interval of abnormal demand fluctuations; based on the probability interval and the confidence parameter, a demand forecast result is generated including a demand forecast value, a forecast confidence interval, and a risk level identifier; Among them, the demand quantity forecast value is a continuous value, which indicates the resource or task request intensity of the system at a specific time point; the demand type forecast result is given in the form of a classification vector, indicating the type of demand occurring at the node, including energy request, maintenance alarm, data upload and task collaboration; the demand time forecast appears in the form of a timestamp or time window indicator, indicating the time range or peak time point when the demand event is expected to occur.

2. The large-model demand forecasting method based on deep learning according to claim 1 is characterized in that: The trusted node set that meets the reliability threshold is screened out based on the data integrity index and transmission stability index of the IoT device, including: Real-time monitoring of the data packet integrity rate and transmission delay jitter of IoT devices to obtain data integrity indicators and transmission stability indicators respectively; Performing weighted calculation based on the data integrity index and the transmission stability index according to preset weight coefficients to establish a device reliability evaluation matrix containing multiple time windows; Based on the device reliability evaluation matrix, the comprehensive trust value of each IoT device is calculated by combining data accuracy, data freshness and data coverage; The comprehensive trust value is compared with the reliability threshold, and the IoT devices whose trust values ​​are greater than the reliability threshold are selected as the trusted node set.

3. The large-model demand forecasting method based on deep learning according to claim 1 is characterized in that: The method of constructing a decentralized network using the geographic coordinate matrix and the communication delay matrix of the trusted node set and adding noise to the original data of the trusted node set to generate noisy training data includes: Obtaining the longitude and latitude coordinates of each IoT device in the trusted node set, establishing a geographic coordinate matrix, and constructing a communication delay matrix by collecting network communication delay data between each IoT device; Calculating the physical distance value between each IoT device based on the geographic coordinate matrix, and calculating the network connection weight value in combination with the communication delay matrix and the physical distance value; Constructing an adjacency matrix based on the network connection weight values, and dynamically adjusting the adjacency matrix according to the time-varying reconnection probability to generate a decentralized network; Device-level Laplace noise and aggregation-level Laplace noise are respectively added to the original data of the trusted node set to generate noisy training data.

4. The large-model demand forecasting method based on deep learning according to claim 3 is characterized in that: The calculating the physical distance value between each IoT device based on the geographic coordinate matrix, and calculating the network connection weight value in combination with the communication delay matrix and the physical distance value, includes: Extract the latitude and longitude values ​​of each IoT device from the geographic coordinate matrix, and calculate the physical distance between the IoT devices using a spherical distance formula combined with the Earth radius constant; The geographical distance influence factor and the communication delay influence factor are set as weight distribution parameters, and the distance standardization parameter and the delay standardization parameter are set as attenuation control variables; Calculating a distance weight component through an exponential decay operation based on the physical distance value and the distance normalization parameter, and calculating a delay weight component through an exponential decay operation based on the delay data in the communication delay matrix and the delay normalization parameter; The distance weight component is weighted according to the geographic distance impact factor to obtain a first connection weight value, the delay weight component is weighted according to the communication delay impact factor to obtain a second connection weight value, and the first connection weight value and the second connection weight value are added to obtain the network connection weight value between each IoT device.

5. The large-model demand forecasting method based on deep learning according to claim 4 is characterized in that: The adding device-level Laplace noise and aggregation-level Laplace noise to the original data of the trusted node set to generate noisy training data includes: Calculate a device-level privacy budget based on the sensitive data ratio and computing power characteristics of each IoT device in the trusted node set, and calculate an aggregate-level privacy budget based on the training rounds and the maximum privacy budget; Generating corresponding Laplace distribution parameters based on the device-level privacy budget value and the global sensitivity parameter, and adding device-level Laplace noise to the original data of each IoT device based on the Laplace distribution parameters to obtain device-level noisy gradient data; Aggregating the device-level noisy gradient data by average according to the number of devices to obtain average aggregated data, and generating aggregate-level Laplace distribution parameters based on the aggregate-level privacy budget value and the aggregate sensitivity parameter; Based on the aggregation-level Laplace distribution parameters, aggregation-level Laplace noise is added to the data after the average aggregation process to generate noisy training data.

6. The large-model demand forecasting method based on deep learning according to claim 1 is characterized in that: The method of performing federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model includes: Based on the adjacency matrix of the decentralized network, a graph convolutional network is constructed to process spatial correlation features, and a long short-term memory network is constructed to process temporal dependency features, thereby obtaining a first spatiotemporal fusion neural network model. Inputting the noise-added training data into the first spatiotemporal fusion neural network model of each node to perform local training calculation to obtain a local gradient vector; Based on the neighbor connection relationship of each node in the decentralized network, the local gradient vector is distributedly aggregated to calculate the gradient difference between neighbor nodes; Aggregation weight coefficients are allocated based on the data quality scores and device reliability assessment results of each IoT device. At the same time, malicious nodes whose gradient differences exceed a preset threshold are detected and excluded from the distributed aggregation process, and a second spatiotemporal fusion neural network model is output.

7. The large-model demand forecasting method based on deep learning according to claim 6 is characterized in that: The distributed aggregation of the local gradient vectors based on the neighbor connection relationship of each node in the decentralized network and the calculation of the gradient difference between neighbor nodes include: Determining a neighbor node set of each IoT device node based on a neighbor connection relationship of each node in the decentralized network; Each IoT device node broadcasts its own local gradient vector to the set of neighboring nodes, and simultaneously receives gradient vector data from neighboring nodes to construct a comparison matrix between the local gradient vector and the neighboring gradient vector; A difference vector between the local gradient vector of each IoT device node and the gradient vector of each neighboring node is calculated based on the comparison matrix, and a gradient difference between the neighboring nodes is generated based on the difference vector.

8. A large-model demand forecasting system based on deep learning, characterized by: A method for implementing a large model demand forecasting method based on deep learning according to any one of claims 1 to 7, comprising: A screening module is used to screen out a set of trusted nodes that meet the reliability threshold based on the data integrity index and transmission stability index of IoT devices; A construction module, configured to construct a decentralized network using the geographic coordinate matrix and the communication delay matrix of the trusted node set, and to add noise to the original data of the trusted node set to generate noisy training data; a federated learning module, configured to perform federated learning on the first spatiotemporal fusion neural network model using the noisy training data in the decentralized network to obtain a second spatiotemporal fusion neural network model; A generation module is used to collect current sensor data and device status information of each IoT device node in the target IoT environment to construct real-time data; input the real-time data into the graph convolution network in the second spatiotemporal fusion neural network model to extract spatial correlation features and generate a spatial feature vector; perform time-series dependency feature calculation on the spatial feature vector through the long short-term memory network in the second spatiotemporal fusion neural network model to generate a time-series feature vector; perform matrix operations on the time-series feature vector and the IoT device state transition matrix to obtain fused spatiotemporal correlation features, and perform nonlinear transformation based on the fused spatiotemporal correlation features to output a multivariate demand forecast value including demand quantity forecast, demand type forecast, and demand time forecast; A Markov state transition chain is constructed based on the historical operating status data of the IoT devices in the trusted node set, and a probability distribution matrix of each IoT device transitioning from a current state to different target states is calculated; the multivariate demand forecast value is decomposed according to the demand quantity dimension, demand type dimension, and demand time dimension to obtain the demand change rate and the demand growth trend coefficient; based on the probability distribution matrix and the demand change rate, a Markov prediction algorithm is used to calculate the device state evolution path in the future time window to obtain the impact weight of the device state change on the demand fluctuation; the demand growth trend coefficient is weighted and corrected according to the impact weight, and a demand fluctuation threshold is introduced for risk assessment to identify the probability interval of abnormal demand fluctuations; based on the probability interval and the confidence parameter, a demand forecast result is generated including a demand forecast value, a forecast confidence interval, and a risk level identifier; Among them, the demand quantity forecast value is a continuous value, which indicates the resource or task request intensity of the system at a specific time point; the demand type forecast result is given in the form of a classification vector, indicating the type of demand occurring at the node, including energy request, maintenance alarm, data upload and task collaboration; the demand time forecast appears in the form of a timestamp or time window indicator, indicating the time range or peak time point when the demand event is expected to occur.

Images