Industrial multi-parameter time series prediction method and system based on large language model

Abstract

The application discloses an industrial multi-parameter time series prediction method and system based on a large language model, which cuts industrial multi-parameter time series data through a sliding window to construct prediction window data, the prediction window data including historical observation values of multiple parameters at continuous time points, each prediction window corresponding to a prediction task; structured information for describing the prediction task is extracted based on the prediction window data, and the structured information at least includes statistical characteristics reflecting the change conditions of the parameters; prompt information is generated through the structured information; the prediction window data is coded into time series features; the correlation between prompt features obtained through the prompt information and the time series features is processed through cross-modal alignment in a parameter granularity to obtain alignment features that fuse numerical information and semantic information; and feature fusion is performed based on the time series features and the alignment features to obtain fusion features for predicting multi-parameter time series at multiple future time steps.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing and artificial intelligence technology, specifically relating to an industrial multi-parameter time series forecasting method and system based on a large language model. Background Technology

[0002] Multi-parameter time series forecasting has extremely high application value in fields such as energy system scheduling, industrial production process monitoring, financial asset quantitative analysis, and intelligent traffic flow guidance. In real-world scenarios, time series typically exhibit high-dimensionality, strong coupling, and non-stationary characteristics. Different parameters not only exhibit evolutionary patterns in the time dimension but also contain intricate spatial correlations and domain semantic associations.

[0003] In existing technologies, multi-parameter time series forecasting schemes are mainly divided into traditional statistical models and deep learning models. Statistical models such as ARIMA and VAR are usually based on the assumption of stationarity and are difficult to capture nonlinear mutation patterns. While deep learning models based on RNN, LSTM and Transformer variants enhance the ability to perceive long-range dependencies, they are essentially still numerical-driven modeling. Their core limitation is that they treat time series as purely digital signals, ignoring the business meaning, physical constraints and macroscopic background information represented by the parameters, which limits the model's generalization ability when dealing with distribution drift or sparse samples.

[0004] To address the aforementioned shortcomings, recent research has attempted to incorporate the commonsense reasoning capabilities of Large Language Models (LLMs) to assist prediction by mapping numerical sequences to a semantic space to enhance feature representation. However, existing LLM-assisted prediction schemes still suffer from significant technical bottlenecks. First, at the modal interaction level, existing technologies mostly employ a one-way logic of "Prompt-guided numerical values," simply treating text prompts as enhancement plugins for numerical features. This design ignores the reverse calibration effect of numerical branches on semantic branches. In large-scale, multi-parameter scenarios, the lack of deep bidirectional intermodal interaction can easily lead to semantic illusions or feature distortion. Second, at the prompt construction level, current schemes heavily rely on fixed templates designed by human experience, lacking an adaptive mechanism that can automatically generate structured semantic descriptions based on the statistical features of window data. This results in extremely high transfer costs for the model when applied across different domains. What is particularly serious is that in actual industrial-grade long-term online prediction tasks, due to the common characteristic of time series alternating between stable and drastic periods, existing technologies perform full large-scale model encoding and inference regardless of whether the data distribution changes. This static processing mechanism, which lacks awareness of error feedback, brings huge computational overhead and response latency, making it difficult to balance prediction accuracy and system operating efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies and achieve adaptive prompt generation, bidirectional depth calibration between modalities, and dynamic caching optimization, this invention adopts the following technical solution:

[0006] An industrial multi-parameter time series forecasting method based on a large language model includes the following process:

[0007] Obtain time series data of multiple industrial parameters and segment them to obtain a set of local data;

[0008] Structured information describing the prediction task is extracted from the local data; the structured information includes trend features of industrial parameter changes, which are used to characterize the overall direction of industrial parameter changes, so that the generated prompt information contains semantic constraints on the future prediction direction.

[0009] Based on the structured information, prompt information for large language models is generated, and different parameter channels in the local data are encoded separately so that each parameter corresponds to an independent feature representation, thereby obtaining local time series features. This provides a basis for subsequent cross-modal alignment at the parameter granularity and solves the problem of insufficient multi-parameter relationship modeling in the prior art.

[0010] Feature extraction is performed on the prompt information, and the obtained prompt features are aligned with the local time series features at the parameter granularity. The bidirectional cross-modal alignment includes aligning the prompt features based on the local time series features, mapping the predicted semantic information to the time series features of the corresponding parameters, and aligning the local time series features based on the prompt features, and adjusting the semantic representation of the prompt features in reverse to obtain aligned features, so as to realize bidirectional interaction between time series numerical information and prompt semantic information.

[0011] Based on the time series features and the alignment features, feature fusion is performed to obtain fused features, so as to solve the problem that semantic noise is easily introduced and the prediction is unstable when semantic information is directly superimposed on numerical features in the prior art.

[0012] Based on the fusion features, prediction results for multi-parameter time series at future time steps are generated.

[0013] Based on the weight information generated during the cross-modal alignment or gated fusion process, the degree of influence of each parameter on the prediction result is output to improve the interpretability of the prediction result.

[0014] Furthermore, the bidirectional cross-modal alignment introduces a topological mask matrix constructed based on the physical spatial connection relationship of sensors in industrial equipment to restrict the semantic information interaction between parameter channels without physical causal correlation. The bidirectional cross-modal alignment includes aligning the cue features based on local time series features under the constraints of the topological mask matrix, mapping the predicted semantic information to the time series features of the corresponding parameters, and aligning the local time series features based on the cue features under the constraints of the topological mask matrix, and adjusting the semantic representation of the cue features in reverse to obtain the aligned features.

[0015] The consistency of the preliminary prediction trend based on the fusion features is checked using preset industrial physical mechanism rules. If the preliminary prediction trend is found to violate the industrial physical mechanism rules, a dynamic penalty mechanism is triggered to reduce the gating weight of the alignment features in the feature fusion process and regenerate the fusion features.

[0016] Based on the final fusion features, prediction results for multi-parameter time series at future time steps are generated.

[0017] This ensures that time-series features only absorb semantic descriptions related to their physical properties, allowing the abstract semantic representation of the large language model to be finely tuned based on the real-time physical correlation of local sensors. This avoids semantic noise interference between parameters without physical correlation, enabling time-series features and prompt features to establish a connection not only at the parameter granularity level but also at the physical logic level, achieving non-uniform sparse alignment and obtaining aligned features that reflect industrial mechanism correlations. Furthermore, by introducing a physical mechanism consistency verification branch after feature fusion generation and before prediction output, the system can quickly cut off interference when the large model produces severe illusions, returning to robust time-series numerical prediction. This effectively isolates the semantic misleading that the language model may bring under abnormal operating conditions, ensuring the safety and robustness of industrial site prediction.

[0018] Furthermore, a sliding prediction window is constructed to slide and divide the time series data of the industrial parameters to obtain prediction window data. The prediction window data includes historical observations of multiple industrial parameters at consecutive time points, where each prediction window data corresponds to a prediction task.

[0019] Furthermore, the structured information includes the time range of the prediction window, the prediction step size, and statistical features reflecting the changes in each parameter. These statistical features include, but are not limited to, trend features within the prediction window's time range, as well as one or more of the mean, standard deviation, and extreme values. Trend features characterize the overall direction of change of industrial parameters within the prediction window, allowing statistical features to directly participate in the prediction guidance of the language model. This ensures that the generated prompts contain semantic constraints on the future prediction direction, thereby enhancing the consistency between the language model's output semantics and the actual changes in the time series. By introducing structured information, the implicit change patterns and trend information in the original numerical sequence are made explicit, providing a clear and calculable input basis for the automatic generation of subsequent prompts. This solves the problem of prompts relying on manual experience in existing technologies. Prompts are automatically generated based on the structured information of the prediction window, and prompt features are cross-modal aligned with time series features at the parameter granularity, allowing the semantic information of the large language model to participate in the prediction modeling process. In addition, by limiting the statistical features of each parameter in the multi-parameter time series within the prediction window, prompts are automatically generated based on dynamically extracted structured fields from the prediction window, avoiding manual design of prompts and improving the adaptability of the prediction task.

[0020] The structured information is used as a field to fill the parameterized template to generate prompt information for large language model processing. The prompt information is generated by "structured fields + template rendering", which makes the prompt construction process have clear input and output rules and avoids the problem of prompts relying on manual design and being difficult to reuse across scenarios in the prior art.

[0021] The historical observations of each parameter within the prediction window are mapped to time series features to characterize the historical change patterns of each parameter within the prediction window.

[0022] Furthermore, based on the correlation between the local time series features and the prompt features, through alignment processing in at least two directions, under the constraint of the correspondence between each parameter in the prediction window and the corresponding semantic fields in the prompt information, cross-modal alignment processing is performed on the local time series features and prompt features at the parameter granularity. This ensures that the semantic information in the prompt features corresponds one-to-one with the local time series features of each parameter, resulting in aligned features that integrate numerical and semantic information. Unlike unidirectional semantic assistance methods, this invention introduces bidirectional cross-modal alignment into the time series prediction scenario and achieves bidirectional interaction between semantic and numerical information at the parameter granularity.

[0023] Furthermore, the bidirectional cross-modal alignment includes a first-direction alignment process and a second-direction alignment process;

[0024] In the first direction alignment process, the local time series features are used as query terms, and the prompt features are used as key terms. The prompt features are aligned based on the local time series features. The local time series features corresponding to the parameter channels are normalized, and the predicted semantic information is mapped to the time series features of the corresponding parameters. The calculation method can be expressed as follows:

[0025]

[0026] in, Based on time series characteristics Mapped to obtain, and Based on prompt features The mapping yields d, which represents the feature dimension.

[0027] In the second-direction alignment process, the prompt feature is used as the query term, and the local time series feature is used as the key value term. The local time series feature is aligned based on the prompt feature, the data corresponding to the local time series feature is normalized, and the semantic representation of the prompt feature is adjusted in reverse to obtain the alignment feature. Its calculation method can be expressed as follows:

[0028]

[0029] in, Based on prompt features Mapped to obtain, and Based on time series characteristics The mapping is obtained.

[0030] Different normalization strategies are adopted for different directions of bidirectional alignment to improve the stability of cross-modal alignment. Through bidirectional cross-modal alignment processing, time series features and cue features are correlated at the parameter granularity to obtain aligned features that integrate numerical and semantic information.

[0031] Furthermore, the prompt information and / or prompt features are stored in the cache unit to reuse the generated prompt information and / or prompt features in the prediction task. The cache unit uses the prediction task-related information as the index key to manage the cached content. The index key includes at least one of the following: the time range corresponding to the prediction window, the prediction step size, and the parameter set identifier, thereby ensuring that the cached prompt information or prompt features have a matching relationship with the current prediction task.

[0032] When a new prediction window arrives, the cache unit is queried based on the index key corresponding to the current prediction window. If the cache unit contains prompt information and / or prompt features that match the current prediction task, the corresponding content is read directly from the cache unit, and the call to the pre-trained language model is skipped. If no match is found in the cache unit, the prompt information is regenerated, the pre-trained language model is called to obtain new prompt features, and these features are written to the cache unit. Through this caching mechanism, repeated calls to the pre-trained language model can be significantly reduced while ensuring the consistency of the prediction logic, thereby reducing computational resource consumption and shortening the prediction response time.

[0033] Furthermore, to prevent cached content from becoming invalid after changes in data distribution, a cache refresh mechanism is introduced. The cached content is refreshed when at least one of the following conditions is met:

[0034] Prediction error exceeds preset threshold: When the actual observation value is obtained and the prediction error is calculated, if the prediction error statistics within the sliding time window exceed the preset threshold, it is determined that the currently cached prompt information and / or prompt features are no longer suitable for the current prediction task.

[0035] The data distribution is determined to have changed based on the statistical features corresponding to the prediction window: when the difference between the extracted statistical features and the statistical features corresponding to the historical prediction windows in the cache exceeds a preset threshold, it is determined that the data distribution has changed.

[0036] When a cache refresh is triggered, you can choose to refresh only the cache of the prompt information or the prompt features, or refresh both the prompt information and the prompt features at the same time, without having to recalculate the time series features or retrain the prediction model. This will further reduce the system's operating overhead while ensuring the accuracy of the prediction.

[0037] By employing caching and refresh mechanisms, the multi-parameter time series prediction method of this invention can adaptively adjust the frequency of calling the large language model according to changes in data distribution in long-term online operation scenarios, balancing prediction accuracy and computational efficiency, and improving the overall practicality and stability of the system.

[0038] Furthermore, the feature fusion employs a gated fusion mechanism; the gate coefficient is calculated based on the local time series features and the aligned features, and its calculation method can be expressed as follows:

[0039]

[0040] Where [· ; ·] denotes feature concatenation operation, σ represents the learnable parameter matrix, and σ represents the Sigmoid function, which is used to constrain the gating coefficients to the range of 0 to 1.

[0041] Based on the gating coefficient, the time series features and alignment features are weighted and combined to generate fused features, the calculation method of which can be expressed as follows:

[0042]

[0043] Here, ⊙ represents element-wise multiplication.

[0044] By employing a gated fusion approach, the intensity of semantic information's involvement in the prediction process can be adaptively adjusted based on the historical variation patterns of different parameters and the correlation between semantic and numerical features. When the semantic information and time-series features are highly consistent, the gate coefficient increases, thereby enhancing the influence of semantic information on the prediction results. Conversely, when semantic information contains biases or noise, the gate coefficient decreases, thus reducing the interference of semantic information on the prediction results. This approach ensures the effective utilization of semantic information while improving the stability and robustness of the prediction results and avoiding semantic noise interference.

[0045] The industrial multi-parameter time series prediction system based on a large language model includes a data input module, a prompt generation module, a time series encoding module, a prompt encoding module, a bidirectional cross-modal alignment module, a feature fusion module, and a prediction output module. It employs the aforementioned industrial multi-parameter time series prediction method based on a large language model, sequentially executing the following steps: acquiring industrial parameter time series data for multiple parameters to generate local data; extracting structured information to generate prompt information; encoding the local data; encoding the prompt information; aligning the prompt features with the encoded local time series features; fusing the time series features with the aligned features; and outputting the prediction result based on the fused features.

[0046] The advantages and beneficial effects of this invention are as follows:

[0047] This invention achieves an effective combination of semantic information and time series features through the combined use of techniques such as prediction window construction, automatic generation of prompts, bidirectional cross-modal alignment, and feature fusion. It improves the accuracy, stability, and interpretability of multi-parameter time series prediction without manual intervention, making it suitable for resource-constrained multi-parameter time series prediction applications. Specifically, this invention constructs a prediction window from multi-parameter time series data and extracts structured information from the prediction window data, such as the time range, prediction step size, and parameter changes, to automatically generate prompts for large language model processing. This eliminates the need for manual design of prompts during time series prediction, avoiding the problems of manual construction and difficulty in adapting to different prediction scenarios in existing technologies. Furthermore, by encoding the multi-parameter time series data corresponding to the prediction window and the automatically generated prompts, time series features and prompt features are obtained. Bidirectional cross-modal alignment is then performed at the parameter granularity, enabling a bidirectional information interaction relationship between time series features and semantic features. This allows the semantic information contained in the language model to effectively participate in time series prediction modeling, improving the insufficient fusion of semantic information and numerical features in existing technologies. Simultaneously, by gating the fusion of bidirectional cross-modal aligned features, the strength of the semantic information's role in the prediction process is dynamically adjusted according to the correlation between time series features and semantic features, thereby reducing the interference of semantic noise on the prediction results and improving the stability and accuracy of the prediction results. Furthermore, by caching the prompt information or prompt features and refreshing the cache when the prediction error increases or the data distribution changes, repeated calls to the large language model in continuous prediction scenarios can be reduced, lowering computational resource consumption and alleviating the problem of high computational overhead during long-term online prediction. Attached Figure Description

[0048] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the structure of the bidirectional cross-modal alignment module in an embodiment of the present invention.

[0051] Figure 4 This is a flowchart illustrating the execution process of the prompt generation module in an embodiment of the present invention.

[0052] Figure 5 This is a schematic diagram of the feature fusion and prediction structure in an embodiment of the present invention. Detailed Implementation

[0053] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] In the field of industrial production process monitoring, large language models, with their powerful semantic understanding and generation capabilities, have shown promising application prospects in complex pattern modeling and prediction tasks. Introducing large language models into time series prediction scenarios helps to characterize the implicit trends and potential patterns in time series from a high-level semantic perspective, thereby improving prediction accuracy. This is especially important in multi-parameter time series prediction, where modeling the complex relationships between different parameters is crucial. However, large language models are typically characterized by large parameter scale and high computational cost. In application scenarios such as multi-parameter time series prediction that require frequent online inference, directly calling large language models will result in significant computational costs and system latency, making it difficult to meet the real-time and resource efficiency requirements of practical applications.

[0055] To reduce the resource consumption of large language models during the prediction process, existing technologies typically optimize them by reducing the frequency of model calls and simplifying the model structure. However, these methods often fail to maintain prediction performance while reducing resource overhead, especially in multi-parameter time series prediction scenarios where there are complex relationships between different parameters. Simply reducing the model's capabilities can easily lead to a decrease in prediction accuracy.

[0056] Therefore, how to fully utilize the semantic modeling capabilities of large language models while ensuring the performance of multi-parameter time series forecasting in industrial production, effectively controlling computational resource consumption, and improving the stability and engineering usability of the forecasting process has become an urgent technical problem to be solved.

[0057] To address the aforementioned problems, this invention proposes an industrial multi-parameter time series prediction method based on a large language model. By introducing a structured prompt information generation mechanism, a bidirectional cross-modal alignment mechanism between time series features and prompt features, a gating fusion mechanism, and a caching and refreshing mechanism for prompt information or prompt features, this method significantly reduces the overhead of repeated calls to the large language model while maintaining prediction accuracy. This improves the practicality of multi-parameter time series prediction in resource-constrained scenarios. Figure 1 As shown, the method of the present invention specifically includes the following steps:

[0058] Step 101: Acquire time series data of multiple parameters in the industrial production process through industrial sensors, and construct a prediction window based on the time series data. The multiple parameters can be different types of parameters obtained from different data acquisition channels.

[0059] First, the data input module acquires different types of industrial production time series data X. Multi-parameter time series data can originate from sensor acquisition systems, operational databases, or real-time monitoring systems, and its data format can be uniformly represented as follows:

[0060]

[0061] Where T represents the length of the time dimension, and M represents the number of parameters. This represents the observed value of the j-th parameter at time point t.

[0062] To address issues such as missing values, dimensional differences, and inconsistent time scales in raw time series data, preprocessing operations are performed on the multi-parameter time series data before constructing the prediction window. These operations include, but are not limited to, missing value imputation, outlier smoothing, and normalization. Normalization can be performed independently for each parameter channel to eliminate the impact of dimensional differences between different parameters on subsequent modeling.

[0063] After preprocessing, the multi-parameter time series data is segmented according to the preset time window length L and sliding step size S to construct the prediction window data. :

[0064]

[0065] By using the above method, continuous time series data is converted into multiple prediction windows with a uniform time length, thereby transforming the long-term series prediction problem into multiple local prediction tasks, reducing the difficulty of model learning and improving prediction stability.

[0066] Step 102: Extracting structured information from the prediction window.

[0067] For the prediction window data obtained in step 101 Structured information describing the prediction task is extracted according to preset feature extraction rules. Structured information includes at least the time range of the prediction window. , +L-1], prediction step size H, and statistical characteristics reflecting the changes in each parameter.

[0068] By introducing structured information This method makes explicit the change patterns and trend information implicit in the original numerical sequence, thereby providing a clear and calculable input basis for the automatic generation of subsequent prompts, solving the problem that prompts in existing technologies rely on human experience for construction.

[0069] In one embodiment, for the j-th parameter, its statistical characteristics are calculated within the prediction window, including:

[0070] Mean:

[0071]

[0072] Standard deviation:

[0073]

[0074] Trend characteristics:

[0075]

[0076] Among them, trend features are used to characterize the overall direction of parameter change within the prediction window, so that the subsequently generated prompts can contain semantic constraints on the future prediction direction, thereby enhancing the consistency between the semantic output of the language model and the actual changes in the time series.

[0077] Step 103: Prompt information generation and time series feature encoding.

[0078] The structured information obtained in step 102 As a field, it is populated into a preset text template, automatically generating prompts for large language model processing. The text template is a parameterized template, which contains placeholder fields for describing the time range, prediction step size, and parameter statistical features.

[0079] By generating prompts using a "structured field + template rendering" approach, the prompt construction process has clear input and output rules, avoiding the problems of prompts relying on manual design and being difficult to reuse across scenarios in existing technologies.

[0080] At the same time, for the prediction window data Encoding is performed to map the historical observations of each parameter within the prediction window to time series features. Its calculation method can be expressed as:

[0081]

[0082] in, This represents a time series coding function. In one embodiment, the time series coding function encodes different parameter channels separately, so that each parameter corresponds to an independent feature representation. This provides a foundation for subsequent cross-modal alignment at the parameter granularity, and solves the problem of insufficient multi-parameter relationship modeling in existing technologies.

[0083] Furthermore, before performing bidirectional cross-modal alignment, this embodiment introduces a parameter topology sensing mechanism for industrial equipment. Specifically, based on the physical spatial connections of sensors in the industrial production system and the sequence of processes (e.g., upstream valve opening and downstream pipeline pressure), a parameter dependency topology graph is constructed. For M parameters in the set of local data, a topological adjacency matrix A∈R is calculated. M×M If the i-th parameter directly affects the j-th parameter in terms of physical mechanism, then matrix element A i,j A high weight is assigned to a topological adjacency matrix A; otherwise, a low weight is assigned. Subsequently, a topology mask matrix, denoted as M, is generated based on the topological adjacency matrix A. topo This topological mask matrix will serve as a priori physical constraint to guide cross-modal alignment between subsequent large language model prompt features and time series features, avoiding semantic noise interference between parameters that are not physically related.

[0084] Step 104: Perform bidirectional cross-modal alignment.

[0085] The time series features obtained in step 103 With the prompt message Hint features obtained by encoding through a pre-trained language model It performs bidirectional cross-modal alignment processing at the parameter granularity to achieve bidirectional interaction between time series numerical information and prompt semantic information.

[0086] Among them, prompt features It can be represented as:

[0087]

[0088] Pooling represents the process of pooling the intermediate representations output by the language model to obtain a fixed-dimensional cue feature representation.

[0089] In one embodiment, bidirectional cross-modal alignment includes a first-direction alignment process and a second-direction alignment process.

[0090] In the first direction alignment process, time series features are used. As a query term, to suggest features As a key-value item, the time-series features that map predicted semantic information to corresponding parameters can be calculated as follows:

[0091]

[0092] in, Based on time series characteristics Mapped to obtain, and Based on prompt features The mapping yields d, which represents the feature dimension.

[0093] In the second-direction alignment process, features are highlighted. As a query term, based on time series features As a key-value item, it performs a reverse adjustment on the semantic representation of the cue features, and its calculation method can be expressed as follows:

[0094]

[0095] in, Based on prompt features Mapped to obtain, and Based on time series characteristics The mapping is obtained.

[0096] Through the above bidirectional cross-modal alignment process, the time series features and cue features are correlated at the parameter granularity, resulting in aligned features that fuse numerical and semantic information.

[0097] In one embodiment, step 104 performs topology-aware bidirectional cross-modal alignment, specifically:

[0098] The time series features obtained in step 103 and the prompt features obtained by encoding the prompt information through a pre-trained language model are subjected to bidirectional cross-modal alignment processing with physical topological constraints at the parameter granularity.

[0099] In this embodiment, to avoid semantic interference between physically unrelated parameters, an industrial parameter topology correlation matrix is ​​first constructed before alignment. Based on the physical location relationship of sensors in industrial equipment or process flow logic (e.g., the frequency of an upstream pump is directly related to the flow rate of the downstream pump), a topology mask matrix M is constructed. topo ∈R M×M , where M is the number of parameter channels. When parameter i and parameter j have no physical association, the mask value is set to negative infinity; otherwise, it is set to 0.

[0100] (1) First-direction alignment processing (numerical to semantic alignment):

[0101] Using time-series features as the query term (Q) and prompt features as the key-value terms (K, V), the predicted semantic information is mapped to the corresponding time-series features. To introduce physical constraints, the calculation method is improved as follows:

[0102]

[0103] The purpose of this step is to limit the scope of semantic information flow using topological masks, ensuring that time series features only absorb semantic descriptions that are physically related to them.

[0104] (2) Second direction alignment processing (semantic to numerical reverse adjustment):

[0105] Using the cue features as the query term (Q) and the time series features as the key-value terms (K, V), the semantic representation of the cue features is adjusted in reverse. The calculation method is improved as follows:

[0106]

[0107] The purpose of this step is to enable the abstract semantic representation of the large language model to be finely tuned based on the real-time physical correlation of local sensors, through M topo Remove irrelevant numerical noise.

[0108] The topological mask matrix M topo It can be automatically generated based on the physical bus topology of industrial equipment, or dynamically constructed by calculating the Granger causality or mutual information between parameters. This ensures that the mask not only includes static physical connections, but also captures dynamic industrial logical relationships.

[0109] Through the bidirectional cross-modal alignment processing with topological masking, time series features and cue features are not only linked at the parameter granularity, but also non-uniform sparse alignment is achieved at the physical logic level, resulting in alignment features that can reflect the correlation of industrial mechanisms.

[0110] Step 105: Feature fusion generation.

[0111] The alignment features obtained in step 104 are fused with the time series features to generate fused features. This addresses the problem in existing technologies where directly superimposing semantic information onto numerical features can easily introduce semantic noise, leading to unstable predictions.

[0112] In one embodiment, the fusion process employs a gated fusion mechanism. Specifically, the gate coefficient G is first calculated based on the time-series features and the features after bidirectional cross-modal alignment. The calculation method can be expressed as follows:

[0113]

[0114] Where [· ; ·] denotes feature concatenation operation, σ represents the learnable parameter matrix, and σ represents the Sigmoid function, which is used to constrain the gating coefficients to the range of 0 to 1.

[0115] Subsequently, based on the gating coefficient G, the time series features and alignment features are weighted and combined to generate fused features. Its calculation method can be expressed as:

[0116]

[0117] Here, ⊙ represents element-wise multiplication.

[0118] By employing the aforementioned gating fusion method, the intensity of semantic information participation in the prediction process can be adaptively adjusted based on the historical variation patterns of different parameters and the correlation between semantic and numerical features. When the semantic information and time-series features are highly consistent, the gating coefficient increases, thereby enhancing the influence of semantic information on the prediction results; conversely, when the semantic information contains biases or noise, the gating coefficient decreases, thereby reducing the interference of semantic information on the prediction results. Thus, while ensuring the effective utilization of semantic information, the stability and robustness of the prediction results are improved.

[0119] Furthermore, to prevent the large language model from generating 'semantic illusions' that deviate from actual physical laws, this embodiment introduces a physical mechanism consistency verification branch after feature fusion generation and before prediction output. Specifically, hard physical mechanism constraints on various parameters in the industrial system are pre-defined, such as temperature and power thresholds set by the law of conservation of energy, or pressure mutation limits in fluid mechanics. Preliminary prediction values ​​are generated based on the fused features, and their physical residuals are calculated. If the preliminary prediction value violates the aforementioned physical mechanism constraints (e.g., the predicted first derivative of a parameter exceeds the maximum response rate of the device hardware), a dynamic penalty mechanism is triggered, forcibly reducing the value of the gating coefficient G mentioned in step 105. The reduction of the gating coefficient G means that the system automatically weakens the weight of the large language model alignment feature Ap→ts, instead relying more on the original time series feature Z. ts The numerical inertia. When the physical residual Δ exceeds the preset threshold, the adjustment logic of the gating coefficient G can be expressed as:

[0120]

[0121] Here, α is the sensitivity factor. This exponential decay ensures that when a large model produces severe illusions, the system can quickly cut off its interference and return to robust time series numerical predictions. This dynamic gating callback mechanism based on physical residuals effectively isolates the semantic misleading that language models may bring under abnormal operating conditions, ensuring the safety and robustness of industrial field predictions.

[0122] Step 106: Generate prediction results.

[0123] The fusion features obtained in step 105 As the input for prediction, the prediction output module generates prediction results for the multi-parameter time series over the next H time steps. Its calculation method can be expressed as follows:

[0124]

[0125] in, This represents the prediction function.

[0126] In one embodiment, to address the problem of high computational overhead and high system latency caused by frequent calls to pre-trained language models during long-term online operation of multi-parameter time series prediction, a caching and refresh mechanism for prompt information or prompt features is introduced in the method of the present invention.

[0127] Specifically, the prompt message in step 103 is as follows. The generation and the suggested features in step 104 After obtaining it, a prompt message will be displayed. and / or prompt features Stored in a cache unit. The cache unit can be located in memory, non-volatile storage media, or a distributed caching system, and is used to reuse the generated prompt information or prompt features in subsequent prediction tasks.

[0128] In one embodiment, the caching unit manages the cached content using prediction task-related information as an index key. The index key includes at least one of the following: the time range corresponding to the prediction window, the prediction step size, and the parameter set identifier, thereby ensuring that the cached prompt information or prompt features have a matching relationship with the current prediction task.

[0129] When a new prediction window arrives, the system queries the cache unit according to the index key corresponding to the current prediction window. If there is a prompt message or prompt feature in the cache unit that matches the current prediction task, the corresponding content is read directly from the cache unit, and the call to the pre-trained language model is skipped. If there is no matching item in the cache unit, the prompt message is regenerated and the pre-trained language model is called to obtain the new prompt feature, and it is written to the cache unit.

[0130] The aforementioned caching mechanism can significantly reduce repeated calls to the pre-trained language model while ensuring the consistency of the prediction logic, thereby reducing computational resource consumption and shortening the prediction response time.

[0131] In one embodiment, to prevent cached content from becoming invalid after changes in data distribution, a cache refresh mechanism is further introduced. The cached content is refreshed when at least one of the following conditions is met:

[0132] (1) Prediction error exceeds the preset threshold: When the system obtains the real observation value and calculates the prediction error, if the prediction error statistics within the sliding time window exceed the preset threshold, it is determined that the currently cached prompt information or prompt features are no longer suitable for the current prediction task.

[0133] (2) Based on the statistical features corresponding to the prediction window, it is determined that the data distribution has changed: When the difference between the statistical features extracted in step 102 and the statistical features corresponding to the historical prediction window in the cache exceeds the preset threshold, it is determined that the data distribution has changed.

[0134] When a cache refresh is triggered, the system can choose to refresh only the cache of the prompt information or the prompt features, or refresh both the prompt information and the prompt features at the same time, without having to recalculate the time series features or retrain the prediction model. This ensures prediction accuracy while further reducing system overhead.

[0135] Through the aforementioned caching and refresh mechanisms, the multi-parameter time series prediction method of this invention can adaptively adjust the calling frequency of the large language model according to changes in data distribution in long-term online operation scenarios, balancing prediction accuracy and computational efficiency, and improving the overall practicality and stability of the system.

[0136] like Figure 2 As shown, the industrial multi-parameter time series prediction system based on a large language model includes: a data input module, a prompt generation module, a time series encoding module, a prompt encoding module, a bidirectional cross-modal alignment module, a feature fusion module, and a prediction output module.

[0137] The data input module is used to acquire multi-parameter time series data and form a prediction window;

[0138] The prompt generation module is used to automatically generate prompt information for large language model processing based on prediction window data;

[0139] The time series encoding module is used to encode multi-parameter time series to obtain time series features;

[0140] The prompt encoding module is used to encode the prompt information to obtain prompt features;

[0141] The bidirectional cross-modal alignment module is used to perform bidirectional cross-modal alignment processing on time series features and cue features;

[0142] The feature fusion module is used to fuse cross-modal aligned features. The feature fusion module adjusts the weights of numerical and semantic information by calculating a gating coefficient G. When sudden anomalies occur in the data, semantic information often lags behind and becomes noise that interferes with prediction. This invention uses a gating mechanism to automatically reduce semantic weights when modal inconsistency is detected, achieving dynamic isolation of semantic noise and enhancing the robustness of the model under extreme conditions.

[0143] The prediction output module is used to output the prediction results of multi-parameter time series based on fused features. Building upon the above prediction process, to further improve the computational efficiency of the system during long-term online operation, in an efficient inference embodiment of this application, the system includes a refresh logic based on prediction error feedback. The system monitors the mean absolute error E of the prediction task in real time. When E is within a preset threshold ϵ for N consecutive periods, the system determines that the current data is in a semantically stable period and directly extracts historical prompt features from the feature cache pool for alignment. When E > ϵ or a data distribution drift (ConceptDrift) is detected, the language model is triggered to re-encode to refresh the cache. This mechanism changes the existing practice of fully calling the large model regardless of whether the data distribution changes. By perceiving the prediction error, it achieves on-demand computation, resulting in a reduction of the system's token consumption and inference time by more than 60% while maintaining accuracy.

[0144] like Figure 3 As shown, the specific structure of the bidirectional cross-modal alignment module includes a time-series feature input terminal, a cue feature input terminal, a first multi-head attention unit, and a second multi-head attention unit. Unlike existing technologies that only use semantics as a numerical supplement, this invention achieves reverse constraint of semantics through the second multi-head attention unit. Existing unidirectional alignment schemes easily lead to a 'semantic illusion' in large language models that deviates from actual numerical trends. This invention uses cue features as Q, time-series features as K and V, and leverages the distribution patterns of real-time observations to forcibly calibrate semantic features. This bidirectional mutual calibration ensures high consistency of cross-modal features at the parameter granularity, resulting in a significant reduction in the bias of prediction results under non-stationary fluctuations.

[0145] Specifically, the time-series feature input receives encoded time-series features, and the cue feature input receives encoded cue features. In the first multi-head attention unit, alignment processing in the first direction is performed. Specifically, using the time-series features as the query term (Q) and the cue features as the key terms (K, V), the predicted semantic information is mapped to the corresponding parameters of the time-series features, ultimately obtaining the aligned time-series feature output. In the second multi-head attention unit, alignment processing in the second direction is performed. Specifically, using the cue features as the query term (Q) and the time-series features as the key terms (K, V), the semantic representation of the cue features is adjusted in reverse based on the time-series features, ultimately obtaining the aligned cue feature output.

[0146] like Figure 4 As shown, the specific execution flow of the prompt generation module includes:

[0147] Step 401 (Receive Window Data): Receive the prediction window data transmitted by the data input module;

[0148] Step 402 (Statistical Indicator Calculation): Based on the prediction window data, calculate the statistical features reflecting the changes in each parameter. In step 402, this application not only extracts the mean and trend, but also further extracts the cross-correlation coefficient reflecting the coupling relationship between parameters and the statistics reflecting the non-stationarity of the sequence. Existing technologies rely on manually designed templates, and the descriptive information is too general. This invention automatically extracts these multi-dimensional structured features and fills them into a preset template. Its function is to provide a higher quality 'numerical profile' for LLM, and the technical effect is to greatly improve the model's adaptability and transfer accuracy across different industry datasets.

[0149] Step 403 (Template Acquisition): Obtain a text template that matches the current prediction task from the preset template library;

[0150] Step 404 (Prompt Text Generation): Use the calculated statistical indicators and time range, and other structured information, as fields, to fill the text template;

[0151] Step 405 (Output Prompt Information): Output the final generated structured prompt information to the prompt encoding module.

[0152] like Figure 5 As shown, the feature fusion and prediction process is implemented through a series of hierarchical structures. First, input data is obtained by receiving aligned time-series features and aligned cue features. Then, the data enters the concatenation layer. In this layer, the time-series features and aligned features are concatenated along the channel dimension or feature dimension, or weighted using a gating mechanism as described in the previous embodiments. The concatenated or fused features are input to the linear projection layer, which adjusts the dimension of the fused features to a dimension space suitable for subsequent models. The adjusted features are input to the large language model base (or the backbone network of a pre-trained language model), utilizing the model's deep semantic reasoning capabilities to capture complex temporal dependencies. The model's output enters the prediction output layer, which is typically composed of fully connected layers or linear regression layers, used to map high-dimensional hidden layer features to values ​​for future time steps, thereby obtaining the final prediction result.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

[0154] To further illustrate the beneficial effects of this invention, we take the multi-parameter prediction of a chemical distillation column system as an example. The distillation column contains strongly coupled parameters such as top temperature, reflux ratio, and feed flow rate. When a non-stationary operating condition occurs with a sudden change in feed flow rate, conventional large-model alignment methods in comparative documents often lack spatial awareness and semantically erroneously align the fluctuation in feed flow rate to the bottom temperature immediately, producing a 'semantic illusion'. However, using this invention, we first construct a topological mask matrix M... topo The system recognizes the physical delay between the sudden change in feed temperature and the temperature at the bottom of the tower, thus preventing the immediate and irrational fusion of the two during cross-modal alignment. Simultaneously, if the large model's suggestive features excessively amplify the temperature prediction, the physical mechanism consistency verification branch will identify a violation of the thermodynamic heating limit and suppress anomalous semantics by lowering the gating coefficient G. Therefore, this invention completely eliminates the uncontrollable risks of black-box large language models in industrial applications while ensuring prediction accuracy. By introducing physical topology masks, the semantic cross-channel interference problem generated by large language models when processing multiple parameters in existing technologies is effectively solved, significantly improving prediction accuracy under strongly coupled conditions. Through physical mechanism consistency verification, the risk of "semantic illusion" that black-box large models may generate in critical industrial decisions is fundamentally eliminated, ensuring the safety of the production system.

Claims

1. An industrial multi-parameter time series forecasting method based on a large language model, characterized by: Obtain time series data of multiple industrial parameters and segment them to obtain a set of local data; Structured information describing the prediction task is extracted from the local data; the structured information includes trend features of industrial parameter changes, which are used to characterize the overall direction of industrial parameter changes, so that the generated prompt information contains semantic constraints on the future prediction direction. Based on the structured information, prompt information for the large language model is generated, and different parameter channels in the local data are encoded respectively so that each parameter corresponds to an independent feature representation, thereby obtaining local time series features; Feature extraction is performed on the prompt information, and the obtained prompt features are aligned with the local time series features at the parameter granularity. The bidirectional cross-modal alignment includes aligning the prompt features based on the local time series features, mapping the predicted semantic information to the time series features of the corresponding parameters, and aligning the local time series features based on the prompt features, and inversely adjusting the semantic representation of the prompt features to obtain the aligned features. Based on the time series features and the alignment features, feature fusion is performed to obtain fused features; Based on the fusion features, prediction results for multi-parameter time series at future time steps are generated.

2. The industrial multi-parameter time series forecasting method based on a large language model according to claim 1, characterized in that: The bidirectional cross-modal alignment introduces a topological mask matrix constructed based on the physical spatial connection relationship of sensors in industrial equipment to restrict the semantic information interaction between parameter channels without physical causal relationship; the bidirectional cross-modal alignment includes aligning cue features based on local time series features in accordance with the topological mask matrix, mapping the predicted semantic information to the time series features of the corresponding parameters, and aligning local time series features based on cue features in accordance with the topological mask matrix, and adjusting the semantic representation of the cue features in reverse to obtain aligned features; The consistency of the preliminary prediction trend based on the fusion features is checked using preset industrial physical mechanism rules. If the preliminary prediction trend is found to violate the industrial physical mechanism rules, a dynamic penalty mechanism is triggered to reduce the gating weight of the alignment features in the feature fusion process and regenerate the fusion features. Based on the final fusion features, prediction results for multi-parameter time series at future time steps are generated.

3. The industrial multi-parameter time series forecasting method based on a large language model according to claim 1, characterized in that: A sliding prediction window is constructed, and the time series data of the industrial parameters is slid-segmented to obtain prediction window data. The prediction window data includes the historical observation values ​​of multiple industrial parameters at continuous time points, and each prediction window data corresponds to a prediction task. The structured information includes the time range of the prediction window, the prediction step size, and statistical features reflecting the changes of each parameter. The statistical features include, but are not limited to, the trend features of the prediction window time range and one or more of the mean, standard deviation, and extreme values. The trend features are used to characterize the overall direction of change of industrial parameters within the prediction window, so that the generated prompt information contains semantic constraints on the future prediction direction.

4. The industrial multi-parameter time series forecasting method based on a large language model according to claim 3, characterized in that: The structured information is used as fields to populate a parameterized template to generate prompts for processing large language models. The historical observations of each parameter within the prediction window are mapped to time series features to characterize the historical change patterns of each parameter within the prediction window.

5. The industrial multi-parameter time series forecasting method based on a large language model according to claim 3, characterized in that: Based on the correlation between the local time series features and the prompt features, through alignment processing in at least two directions, under the constraint of the correspondence between each parameter in the prediction window and the corresponding semantic field in the prompt information, cross-modal alignment processing is performed on the local time series features and prompt features at the parameter granularity, so that the semantic information in the prompt features corresponds one-to-one with the local time series features of each parameter, resulting in aligned features that integrate numerical information and semantic information.

6. The industrial multi-parameter time series forecasting method based on a large language model according to claim 5, characterized in that: The bidirectional cross-modal alignment includes a first-direction alignment process and a second-direction alignment process; In the first direction alignment process, the local time series features are used as query items and the prompt features are used as key items. The prompt features are aligned based on the local time series features. The local time series features corresponding to the parameter channels are normalized, and the predicted semantic information is mapped to the time series features of the corresponding parameters. In the second direction alignment process, the prompt feature is used as the query term, the local time series feature is used as the key value term, the local time series feature is aligned based on the prompt feature, the data corresponding to the local time series feature is normalized, and the semantic representation of the prompt feature is adjusted in reverse to obtain the alignment feature.

7. The industrial multi-parameter time series forecasting method based on a large language model according to claim 3, characterized in that: The prompt information and / or prompt features are stored in the cache unit to reuse the generated prompt information and / or prompt features in the prediction task. The cache unit uses the prediction task-related information as the index key to manage the cached content. The index key includes at least one of the following: the time range corresponding to the prediction window, the prediction step size, and the parameter set identifier. When a new prediction window arrives, the cache unit is queried according to the index key corresponding to the current prediction window. If there is a prompt message and / or prompt feature in the cache unit that matches the current prediction task, the corresponding content is read directly from the cache unit, and the call to the pre-trained language model is skipped. If there is no matching item in the cache unit, the prompt message is regenerated and the pre-trained language model is called to obtain new prompt features, which are then written to the cache unit.

8. The industrial multi-parameter time series forecasting method based on a large language model according to claim 7, characterized in that: A cache refresh mechanism is introduced, which refreshes the cached content when at least one of the following conditions is met: Prediction error exceeds preset threshold: When the actual observation value is obtained and the prediction error is calculated, if the prediction error statistics within the sliding time window exceed the preset threshold, it is determined that the currently cached prompt information and / or prompt features are no longer suitable for the current prediction task. The data distribution is determined to have changed based on the statistical features corresponding to the prediction window: when the difference between the extracted statistical features and the statistical features corresponding to the historical prediction windows in the cache exceeds a preset threshold, it is determined that the data distribution has changed. When a cache refresh is triggered, you can choose to refresh only the cache of the prompt message or the prompt feature, or refresh both the prompt message and the prompt feature at the same time. By employing caching and refresh mechanisms, the multi-parameter time series prediction method of this invention can adaptively adjust the frequency of calling the large language model according to changes in data distribution in long-term online operation scenarios, balancing prediction accuracy and computational efficiency, and improving the overall practicality and stability of the system.

9. The industrial multi-parameter time series forecasting method based on a large language model according to claim 1, characterized in that: The feature fusion employs a gated fusion mechanism; the gate coefficient is calculated based on the local time series features and the aligned features. Based on the gating coefficient, the time series features and alignment features are weighted and combined to generate fused features.

10. An industrial multi-parameter time series forecasting system based on a large language model, comprising a data input module, a prompt generation module, a time series encoding module, a prompt encoding module, a bidirectional cross-modal alignment module, a feature fusion module, and a prediction output module, characterized in that: The method for industrial multi-parameter time series prediction based on a large language model, as described in any one of claims 1 to 9, sequentially performs the following steps: acquiring industrial parameter time series data of multiple parameters to generate local data; extracting structured information to generate prompt information; encoding the local data; encoding the prompt information; aligning the prompt features with the encoded local time series features; fusing the time series features with the aligned features; and outputting the prediction result based on the fused features.

Images