Intelligent prediction and dynamic optimization method for power consumption of optical module based on deep learning

Abstract

The application provides a deep learning-based intelligent prediction and dynamic optimization method for optical module power consumption, relates to the technical field of optical communication, and comprises the following steps: spatiotemporal decoupling of multi-dimensional running data of an optical module to separate a time-domain characteristic sequence and a spatial-domain characteristic atlas; identification of a key time window through an attention mechanism and generation of a power consumption sensitive characteristic vector to establish a causal dependence of the power consumption sensitive characteristic vector on future power consumption; fusion of the spatiotemporal characteristics to obtain a fused characteristic representation, and further deduction of a power consumption prediction value and a fluctuation interval; and finally mapping into a constraint surface in a power consumption-performance joint optimization space, searching for an optimal working point through adaptive search, and generating an optimization control strategy. The application realizes accurate prediction and dynamic optimization of the power consumption of the optical module, and improves energy efficiency and system stability.

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to a method for intelligent prediction and dynamic optimization of optical module power consumption based on deep learning. Background Technology

[0002] As a core component in optical communication networks, the power consumption management of optical modules is crucial for improving network energy efficiency and operational stability. With the continuous expansion of data centers and 5G networks, the number and deployment density of optical modules are increasing dramatically, and their power consumption has become a significant component of network operating costs and carbon emissions. Therefore, accurate prediction and dynamic optimization of optical module power consumption are key technical aspects for achieving green communication networks.

[0003] Currently, conventional practices in this technology primarily rely on statistical models or simplified physical models based on historical data for power consumption estimation. A common approach is to use time series analysis methods, such as autoregressive models or exponential smoothing, to predict future trends based on past power consumption data sequences of optical modules. These methods focus on continuous changes over time, but typically treat each optical module as an independent entity, ignoring the potential impact of spatial correlations between different modules in the network topology due to connectivity and data flow interactions on power consumption. Another common approach is to use the typical power consumption curve calibrated at the module's factory, combined with a few key operating parameters monitored in real time (such as temperature and bias current), to perform linear interpolation or table lookups to estimate the current power consumption state. While this method is simple and direct, it relies on static, pre-set model parameters and struggles to capture the power consumption fluctuations caused by the nonlinear coupling between multiple internal physical parameters (such as laser drive current, receiver sensitivity, and modulation format) of the optical module under complex dynamic workloads.

[0004] Models that treat time and space characteristics separately cannot effectively characterize the complex mechanism by which optical module power consumption in real network environments is simultaneously affected by its own historical state evolution and the collaborative operation of neighboring modules. This leads to a decrease in the accuracy of prediction models under topology changes or sudden traffic scenarios, making it difficult to reflect the dynamic correlation of power consumption at the system level. Estimation methods based on static parameters or simple linear relationships are insufficient in characterizing the nonlinear dynamics of operating characteristics. The power consumption response of an optical module is a complex process involving the interaction of multiple physical domains such as electricity, optics, and thermal, and its dynamic evolution is highly nonlinear. Existing methods cannot deeply analyze the profound causal relationship between different operating characteristics and power consumption fluctuations, especially in accurately quantifying the range of power consumption fluctuations caused by future uncertainties, resulting in prediction results that are often biased and have limited reliability. These shortcomings limit prediction accuracy, which in turn leads to a lack of accurate decision-making basis for subsequent power consumption optimization strategies, making it difficult to achieve optimal energy efficiency while ensuring transmission performance. Summary of the Invention

[0005] This invention provides a method for intelligent prediction and dynamic optimization of optical module power consumption based on deep learning, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for intelligent prediction and dynamic optimization of optical module power consumption based on deep learning, comprising: Spatiotemporal decoupling processing is performed on the multidimensional operational feature data of the optical module to separate and represent the dynamic evolution features in the time dimension and the topological correlation features in the spatial dimension, thereby obtaining the temporal feature sequence and the spatial feature map. The correlation strength of features at different times in the time-domain feature sequence is quantified by an attention weighting mechanism, the key time window with the highest correlation strength with power consumption changes is identified, the features in the key time window are deeply embedded and transformed to generate a power consumption sensitive feature vector, and the causal dependency between the key time window and the future power consumption evolution path is established. The power-sensitive feature vector and the spatial feature map are fused across modes, and the spatiotemporal coupling relationship is captured by constructing a feature interaction matrix to obtain the fused feature representation. Based on the fusion feature representation and the causal dependency, the power consumption prediction value and power consumption fluctuation range are obtained step by step through multi-level nonlinear transformation. A power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multi-dimensional constraint surface. The optimal operating point is found on the multi-dimensional constraint surface by a gradient-guided adaptive search method. Based on the optimal operating point, an optimized control strategy containing operating mode switching instructions and transmission parameter adjustment instructions is generated.

[0007] Spatiotemporal decoupling processing is performed on the multidimensional operational characteristic data of the optical module, separating and representing the dynamic evolution characteristics in the temporal dimension and the topological correlation characteristics in the spatial dimension, resulting in a temporal feature sequence and a spatial feature map, including: The multidimensional operational feature data is sequentially arranged according to the time axis, and the feature change trajectory within a continuous time period is extracted to form a dynamic evolution feature in the time dimension. A time-domain direction vector is constructed based on the change gradient. The multidimensional operational feature data is used to perform topological modeling based on the physical connection relationship and signal transmission path between the internal components of the optical module, extract the spatial association pattern between the components, form the topological association features in the spatial dimension, and construct the spatial direction vector based on the transmission intensity. The time-domain direction vector and the spatial-domain direction vector are orthogonalized to obtain the time-domain basis vector and the spatial-domain basis vector, which are then combined to form a time-space separation transformation matrix. The dynamic evolution features of the time dimension are multiplied by the time-domain basis vectors to obtain the time-domain projection coefficients. The time-domain projection features are then reconstructed in the time-domain subspace based on the time-domain projection coefficients. The topological correlation features of the spatial dimension are multiplied by the spatial basis vectors to obtain the spatial projection coefficients. The spatial projection features are then reconstructed in the spatial subspace based on the spatial projection coefficients. Based on the temporal projection features, the temporal feature sequence is organized in chronological order; based on the spatial projection features, the spatial feature map is constructed according to the topological connection relationship.

[0008] Deep embedding transformation is performed on the features within the key time window to generate power-sensitive feature vectors. Simultaneously, the causal dependency between the key time window and the future power evolution path is established, including: Extract all time features within the key time window and arrange them in chronological order to form a window feature sequence; The window feature sequence is subjected to multi-layer nonlinear mapping transformation. During the transformation process, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimension layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to the power consumption sensitive feature space to obtain the power consumption sensitive feature vector. Obtain historical power consumption measurement data corresponding to the key time window, as well as power consumption measurement data in the future time period after the key time window, and construct a time-aligned power consumption evolution sequence. Calculate the covariance matrix between the power-sensitive feature vector and the future power change in the power evolution sequence. Based on the covariance matrix, identify the degree of influence of each feature component in the power-sensitive feature vector on the future power evolution path. Construct a causal dependency mapping relationship from the feature components to the future power change. Quantify the causal dependency relationship between the key time window and the future power evolution path based on the causal dependency mapping relationship.

[0009] The window feature sequence is subjected to a multi-layer nonlinear mapping transformation. During the transformation process, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimension layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to a power-sensitive feature space to obtain a power-sensitive feature vector, including: Obtain the initial feature dimension and feature value distribution of the window feature sequence, and determine the input features of the window feature sequence; A multi-layered cascaded nonlinear mapping layer is constructed. In each nonlinear mapping layer, a feature transformation unit and a dimension compression unit are set. The feature transformation unit extracts high-order feature representations by performing nonlinear activation operations on the input features. The dimension compression unit reduces feature redundancy by performing dimension reduction operations on the high-order feature representations. A power consumption correlation measurement mechanism is introduced in each nonlinear mapping layer to calculate the correlation coefficient between the higher-order feature representation and the historical power consumption fluctuation pattern. Based on the correlation coefficient, the weights of each feature component in the higher-order feature representation are adjusted to enhance the numerical weights of feature components that are strongly correlated with power consumption fluctuations and suppress the numerical weights of feature components that are weakly correlated with power consumption fluctuations. The window feature sequence is sequentially input into the multi-layer cascaded nonlinear mapping layer for layer-by-layer transformation. In the first nonlinear mapping layer, the window feature sequence is transformed into a first-layer high-order feature representation. In the second nonlinear mapping layer, the first-layer high-order feature representation is transformed into a second-layer high-order feature representation. This process is continued until the last nonlinear mapping layer outputs the last-layer high-order feature representation, which serves as the power-sensitive feature vector in the power-sensitive feature space.

[0010] The power-sensitive feature vector and the spatial feature map are fused across modalities. By constructing a feature interaction matrix to capture the spatiotemporal coupling relationship, the fused feature representation is obtained, including: The power-sensitive feature vector is decomposed in the time domain to obtain time-domain sub-feature vectors at multiple time scales, and the spatial feature map is decomposed in the spatial domain to obtain spatial sub-feature maps at multiple spatial scales. A multi-scale feature interaction matrix is ​​constructed for the temporal sub-feature vectors at each time scale and the spatial sub-feature maps at each spatial scale. A mapping relationship between the time scale and the spatial scale is established in the multi-scale feature interaction matrix, and the cross-scale interaction strength between the temporal sub-feature vectors at different time scales and the spatial sub-feature maps at different spatial scales is calculated. An adaptive attention weighting mechanism is constructed, and dynamic attention weights are assigned to the temporal sub-feature vectors of each time scale and the spatial sub-feature maps of each spatial scale according to the cross-scale interaction intensity. The dynamic attention weights are proportional to the strength of the spatiotemporal coupling relationship. Based on the dynamic attention weights, the temporal sub-feature vectors at each time scale are weighted and summed to obtain attention-enhanced temporal features, and the spatial sub-feature maps at each spatial scale are weighted and summed to obtain attention-enhanced spatial features. The attention-enhanced temporal features and the attention-enhanced spatial features are concatenated to generate a fused feature representation.

[0011] Based on the fused feature representation and the causal dependency, the power consumption prediction value and power consumption fluctuation range are obtained step by step through multi-level nonlinear transformation, including: The causal constraints of the fusion feature are generated by weighting and adjusting each feature component represented by the causal relationship according to the causal weight in the causal dependency relationship. A multi-level nonlinear extrapolation network is constructed, which includes multiple serial extrapolation layers. Each extrapolation layer performs nonlinear transformation on the input features and gradually extracts the power consumption evolution trend features. The last extrapolation layer outputs an extrapolation feature vector containing power consumption center trend information and power consumption uncertainty information. The inferred feature vector is decoupled by branching to construct a power consumption prediction branch and a fluctuation range estimation branch. The power consumption prediction branch performs nonlinear mapping on the power consumption center trend information in the inferred feature vector to obtain the power consumption prediction value. The fluctuation range estimation branch performs probability distribution modeling on the power consumption uncertainty information in the inferred feature vector to obtain the upper and lower bounds of the power consumption fluctuation range. The causal constraint fusion features are input into the multi-level nonlinear inference network for layer-by-layer nonlinear transformation, and the inference feature vector is output. The inferred feature vector is input into the power consumption prediction branch and the fluctuation range estimation branch respectively. The power consumption prediction value is obtained through the power consumption prediction branch, and the power consumption fluctuation range is obtained through the fluctuation range estimation branch.

[0012] A power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multidimensional constraint surface. The optimal operating point is then found on this multidimensional constraint surface using a gradient-guided adaptive search method, including: Collect the power consumption status parameter set and performance index parameter set of the optical module, map the power consumption status parameter set to the power consumption dimension coordinate of the power consumption-performance joint optimization space, map the performance index parameter set to the performance dimension coordinate of the power consumption-performance joint optimization space, determine the power consumption constraint boundary in the power consumption-performance joint optimization space based on the power consumption prediction value and the power consumption fluctuation range, and determine the performance constraint boundary based on the performance requirements of the optical module; The predicted power consumption value is used as the power consumption center reference point, and the upper and lower bounds of the power consumption fluctuation range are used as the power consumption deviation range. The multidimensional constraint surface is constructed in the power consumption-performance joint optimization space by combining the power consumption constraint boundary and the performance constraint boundary. A gradient-guided adaptive search method is constructed. The gradient direction of the power consumption-performance joint objective function is calculated on the multidimensional constrained surface. The power consumption-performance joint objective function comprehensively evaluates power efficiency and performance satisfaction. The search path is determined based on the gradient direction and the step size is adjusted adaptively. When the power consumption-performance joint objective function converges, the optimal operating point is determined.

[0013] A second aspect of the present invention provides a deep learning-based intelligent prediction and dynamic optimization system for optical module power consumption, comprising: The spatiotemporal decoupling unit is used to perform spatiotemporal decoupling processing on the multidimensional operational feature data of the optical module, separating and representing the dynamic evolution features in the time dimension and the topological correlation features in the spatial dimension to obtain the temporal feature sequence and the spatial feature map. The feature weighting unit is used to quantify the correlation strength of features at different times in the time-domain feature sequence through an attention weighting mechanism, identify the key time window with the highest correlation strength with power consumption changes, perform deep embedding transformation on the features in the key time window to generate a power consumption sensitive feature vector, and establish a causal dependency relationship between the key time window and the future power consumption evolution path. A cross-modal fusion unit is used to perform cross-modal feature fusion between the power-sensitive feature vector and the spatial feature map, and to capture the spatiotemporal coupling relationship by constructing a feature interaction matrix to obtain the fused feature representation; The power consumption extrapolation unit is used to gradually extrapolate the predicted power consumption value and the power consumption fluctuation range based on the fused feature representation and the causal dependency relationship through multi-level nonlinear transformation. An optimized control unit is used to construct a power consumption-performance joint optimization space. In the power consumption-performance joint optimization space, the predicted power consumption value and the power consumption fluctuation range are mapped to a multi-dimensional constraint surface. An optimal operating point is found on the multi-dimensional constraint surface through a gradient-guided adaptive search method. Based on the optimal operating point, an optimized control strategy containing operating mode switching instructions and transmission parameter adjustment instructions is generated.

[0014] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] The beneficial effects of this application are as follows: This method enables accurate prediction and dynamic optimization of optical module power consumption, significantly improving the energy efficiency and reliability of optical communication networks. By performing spatiotemporal decoupling processing on the multidimensional operational characteristics of optical modules, the temporal dynamic evolution characteristics and spatial topological correlation characteristics are separated and represented, effectively removing redundancy and noise from the original data. This provides subsequent analysis with a clearer structure and higher information density of temporal feature sequences and spatial feature maps, laying a data foundation for accurate modeling.

[0017] By quantifying the correlation strength of temporal feature sequences using an attention-weighted mechanism, we can automatically identify and focus on the key time windows most relevant to power consumption changes, avoiding the limitations of traditional methods that average data across all time periods. Deep embedding transformation of the key window features generates highly condensed power-sensitive feature vectors. Simultaneously, the established causal dependencies reveal the driving mechanism of historical key states on future power consumption evolution, enabling the prediction model to not only know what happens, but also why, significantly improving the logic and interpretability of the predictions.

[0018] By fusing power-sensitive feature vectors with spatial feature maps across modes and constructing a feature interaction matrix to capture spatiotemporal coupling relationships, the problem of mutual interference between temporal dynamics and spatial correlations is solved. The resulting fused feature representation comprehensively characterizes the overall operating state of the optical module in complex network environments, enabling the prediction model to simultaneously perceive its own historical trajectory and the influence of the surrounding topology, thereby significantly improving the prediction robustness and accuracy in dynamically changing scenarios.

[0019] Based on fusion characteristics and causal dependencies, and through multi-level nonlinear transformations, accurate power consumption predictions and their fluctuation ranges can be output simultaneously. This combination of point estimation and interval estimation not only provides the expected level of future power consumption but also quantifies its uncertainty range, providing crucial information for risk assessment and robust decision-making, and enhancing the system's ability to cope with sudden anomalies and uncertainties.

[0020] A joint power consumption-performance optimization space is constructed, and the prediction results are mapped to a multidimensional constraint surface, realizing the mathematical representation and visualization of the optimization objective. A gradient-guided adaptive search method is employed to efficiently and accurately locate the globally optimal or near-optimal operating point on the complex constraint surface. Based on this, the generated optimization control strategy can dynamically coordinate operating mode switching and transmission parameter adjustment, minimizing or allocating power consumption on demand while strictly ensuring communication performance indicators. This achieves the optimal balance between energy efficiency and performance, extending equipment lifespan and reducing operating costs. Attached Figure Description

[0021] Figure 1This is a flowchart illustrating the intelligent prediction and dynamic optimization method for optical module power consumption based on deep learning, as described in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the process of generating power-sensitive feature vectors and constructing causal dependencies. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0024] Figure 1 This is a flowchart illustrating the intelligent prediction and dynamic optimization method for optical module power consumption based on deep learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Spatiotemporal decoupling processing is performed on the multidimensional operational feature data of the optical module to separate and represent the dynamic evolution features in the time dimension and the topological correlation features in the spatial dimension, thereby obtaining the temporal feature sequence and the spatial feature map. The correlation strength of features at different times in the time-domain feature sequence is quantified by an attention weighting mechanism, the key time window with the highest correlation strength with power consumption changes is identified, the features in the key time window are deeply embedded and transformed to generate a power consumption sensitive feature vector, and the causal dependency between the key time window and the future power consumption evolution path is established. The power-sensitive feature vector and the spatial feature map are fused across modes, and the spatiotemporal coupling relationship is captured by constructing a feature interaction matrix to obtain the fused feature representation. Based on the fusion feature representation and the causal dependency, the power consumption prediction value and power consumption fluctuation range are obtained step by step through multi-level nonlinear transformation. A power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multi-dimensional constraint surface. The optimal operating point is found on the multi-dimensional constraint surface by a gradient-guided adaptive search method. Based on the optimal operating point, an optimized control strategy containing operating mode switching instructions and transmission parameter adjustment instructions is generated.

[0025] In one optional implementation, the multidimensional operational characteristic data of the optical module undergoes spatiotemporal decoupling processing, separating and representing the dynamic evolution characteristics in the time dimension and the topological correlation characteristics in the spatial dimension, resulting in a temporal feature sequence and a spatial feature map, including: The multidimensional operational feature data is sequentially arranged according to the time axis, and the feature change trajectory within a continuous time period is extracted to form a dynamic evolution feature in the time dimension. A time-domain direction vector is constructed based on the change gradient. The multidimensional operational feature data is used to perform topological modeling based on the physical connection relationship and signal transmission path between the internal components of the optical module, extract the spatial association pattern between the components, form the topological association features in the spatial dimension, and construct the spatial direction vector based on the transmission intensity. The time-domain direction vector and the spatial-domain direction vector are orthogonalized to obtain the time-domain basis vector and the spatial-domain basis vector, which are then combined to form a time-space separation transformation matrix. The dynamic evolution features of the time dimension are multiplied by the time-domain basis vectors to obtain the time-domain projection coefficients. The time-domain projection features are then reconstructed in the time-domain subspace based on the time-domain projection coefficients. The topological correlation features of the spatial dimension are multiplied by the spatial basis vectors to obtain the spatial projection coefficients. The spatial projection features are then reconstructed in the spatial subspace based on the spatial projection coefficients. Based on the time-domain projection features, the time-domain feature sequence is organized in chronological order. Based on the spatial projection features, the spatial feature map is constructed according to the topological connection relationship.

[0026] The multidimensional operational characteristic data of the optical module includes real-time acquired physical measurements such as current, voltage, temperature, optical power, and bit error rate, with a sampling frequency of 100 times per second. These characteristic data are serialized and arranged according to timestamps to form a time-series matrix of length T, where each time step contains N feature dimensions. The gradient of change is obtained by calculating the difference between feature values ​​at adjacent time steps; features with gradient values ​​greater than a set threshold of 0.05 are marked as significantly changing features. These gradient values ​​are then organized into a time-domain direction vector according to the feature dimensions. This vector represents the evolution trend of each feature over time.

[0027] A topology graph is constructed based on the physical connections of components such as lasers, modulators, detectors, and driver circuits within the optical module. Nodes in the topology graph represent components, and edge weights represent signal transmission strength. Coupling coefficients between components are extracted by analyzing the transmission paths of electrical and optical signals. A directed connection edge is established when the signal transmission strength from component i to component j exceeds 0.3. The coupling coefficients of all components are arranged in topological order to form a spatial direction vector. , where M is the total number of components.

[0028] Gram-Schmidt orthogonalization is performed on the time-domain direction vector and the spatial-domain direction vector. The time-domain direction vector is then normalized to obtain the time-domain basis vector e. t Subtracting the projection component of the spatial direction vector onto the time-domain basis vector and normalizing the residual vector yields the spatial basis vector e. s Verify that the absolute value of the inner product of two basis vectors is less than 10. -6 Confirm that the orthogonality requirement is met. Combine the time-domain basis vectors and spatial basis vectors column-wise to construct the space-time separation transformation matrix P = [e t e s ].

[0029] The dynamic evolution characteristics in the time dimension are multiplied by the time-domain basis vectors using the inner product operation, and the calculation formula is as follows: , where f t Let α be the feature vector at a certain time. t These are the time-domain projection coefficients. After calculating the time-domain projection coefficient sequence for all time points, it is then processed by f. t，proj = α t e t The temporal projection features are reconstructed in the temporal subspace. These projection features retain the main components of the original features in the direction of temporal evolution and filter out time-independent noise components.

[0030] The spatial projection coefficients are obtained by performing an inner product operation between the topological correlation features of the spatial dimension and the spatial basis vectors. This is achieved through f... s，proj = α s e s Spatial projection features are obtained by reconstructing the spatial subspace. These projection features highlight the spatial coupling relationships between components and suppress interference caused by temporal fluctuations.

[0031] Arrange the temporal projection features at each moment in chronological order to form a structure of length T and dimension D. t The temporal feature sequence. This sequence is represented in matrix form as follows: Each row corresponds to the projection feature at a given time. The spatial projection features of each component are organized according to the node numbers and connections in the topology graph to construct a spatial feature map G represented by an adjacency matrix. s The feature dimension of nodes in the graph is D. s The edge features are the correlation measures of the projection coefficients.

[0032] In one optional implementation, deep embedding transformation is performed on the features within the key time window to generate a power-sensitive feature vector, and the causal dependency between the key time window and the future power evolution path is established, including: Extract all time features within the key time window and arrange them in chronological order to form a window feature sequence; The window feature sequence is subjected to multi-layer nonlinear mapping transformation. During the transformation process, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimension layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to the power consumption sensitive feature space to obtain the power consumption sensitive feature vector. Obtain historical power consumption measurement data corresponding to the key time window, as well as power consumption measurement data in the future time period after the key time window, and construct a time-aligned power consumption evolution sequence. Calculate the covariance matrix between the power-sensitive feature vector and the future power change in the power evolution sequence. Based on the covariance matrix, identify the degree of influence of each feature component in the power-sensitive feature vector on the future power evolution path. Construct a causal dependency mapping relationship from the feature components to the future power change. Quantify the causal dependency relationship between the key time window and the future power evolution path based on the causal dependency mapping relationship.

[0033] like Figure 2 As shown, the method includes: An attention query vector is constructed, initialized by analyzing the feature patterns of historical power consumption mutation samples. For each time step in the time-domain feature sequence, the similarity between the feature vector and the query vector is calculated to obtain the initial attention score. The initial attention scores for all time steps are normalized to obtain the attention weight for each time step; a larger weight indicates a stronger correlation between the feature and power consumption change at that time step.

[0034] A fixed-length time window is defined, and a sliding window approach is used to traverse the entire time-domain feature sequence. For each window position, the cumulative sum of attention weights at all times within that window is calculated; the cumulative sum characterizes the overall correlation strength between that window and power consumption changes. Several windows with the largest cumulative sums are selected as candidate key time windows. To avoid excessive overlap between windows, a non-maximum suppression strategy is adopted: when the number of overlapping times between two candidate windows exceeds half the window length, the window with the larger cumulative sum is retained, and the window with the smaller cumulative sum is discarded.

[0035] After nonmaximum suppression, several key time windows are finally determined, each covering multiple consecutive moments.

[0036] After obtaining the critical time window, the operational characteristics contained within it need to be converted into a compact representation that directly reflects the power consumption variation. Following the chronological order of the sampling timestamps, the feature vectors for each moment within the critical time window are read sequentially and arranged chronologically into an ordered window feature sequence. This sequence preserves the temporal evolution information within the window, providing structured input for subsequent deep transformations.

[0037] For window feature sequences, a nonlinear mapping network consisting of three to five fully connected layers is constructed. The number of neurons in each layer decreases sequentially. The number of neurons in the first layer is set to 1.5 to 2 times the original feature dimension, and the number of neurons in subsequent layers decreases layer by layer with a decay coefficient of 0.6 to 0.8. The output dimension of the last layer is set to 16 to 64 dimensions. Modified linear unit activation functions are used between layers, and batch normalization is introduced between the second and third layers to stabilize the training process. Through the layer-by-layer transformation of this multi-layer network, redundant information unrelated to power consumption in the original feature sequence is gradually filtered out, while abstract patterns that can indicate power consumption fluctuation trends are enhanced and preserved in the output power-sensitive feature vector.

[0038] When establishing causal dependencies, the measured power consumption value sequence corresponding to the key time window is extracted from the historical database. Simultaneously, the measured power consumption value sequence for the consecutive 5 to 30 minutes following the end of the window is obtained. The power consumption sequence within the window and the power consumption sequence after the window are aligned and concatenated on the time axis to form a complete power consumption evolution sequence. The change in power consumption value in future time periods relative to the power consumption value at the end of the window is calculated in this power consumption evolution sequence to obtain the future power consumption change vector.

[0039] Each dimension component of the power-sensitive feature vector is paired with the vector of future power consumption changes to form a sample pair, and the covariance matrix between them is calculated. The element in the i-th row and j-th column of the covariance matrix represents the linear correlation strength between the i-th component of the power-sensitive feature vector and the power consumption change at the j-th future time. Each row of the covariance matrix is ​​normalized to obtain the contribution weight distribution of each feature component to the power consumption change at each future time. Components with contribution weights greater than a set threshold of 0.3 to 0.5 are marked as strong causal features, and components with weights between 0.1 and 0.3 are marked as weak causal features. By storing the causal label of each feature component and its corresponding weight vector, a causal dependency mapping relationship from the power-sensitive feature vector to the future power consumption evolution path is constructed. This mapping relationship clearly indicates how the feature state within the key time window affects the direction and magnitude of subsequent power consumption evolution.

[0040] In one optional implementation, the window feature sequence undergoes a multi-layer nonlinear mapping transformation. During the transformation, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimensions layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to a power-sensitive feature space to obtain a power-sensitive feature vector, including: Obtain the initial feature dimension and feature value distribution of the window feature sequence, and determine the input features of the window feature sequence; A multi-layered cascaded nonlinear mapping layer is constructed. In each nonlinear mapping layer, a feature transformation unit and a dimension compression unit are set. The feature transformation unit extracts high-order feature representations by performing nonlinear activation operations on the input features. The dimension compression unit reduces feature redundancy by performing dimension reduction operations on the high-order feature representations. A power consumption correlation measurement mechanism is introduced in each nonlinear mapping layer to calculate the correlation coefficient between the higher-order feature representation and the historical power consumption fluctuation pattern. Based on the correlation coefficient, the weights of each feature component in the higher-order feature representation are adjusted to enhance the numerical weights of feature components that are strongly correlated with power consumption fluctuations and suppress the numerical weights of feature components that are weakly correlated with power consumption fluctuations. The window feature sequence is sequentially input into the multi-layer cascaded nonlinear mapping layer for layer-by-layer transformation. In the first nonlinear mapping layer, the window feature sequence is transformed into a first-layer high-order feature representation. In the second nonlinear mapping layer, the first-layer high-order feature representation is transformed into a second-layer high-order feature representation. This process is continued until the last nonlinear mapping layer outputs the last-layer high-order feature representation, which serves as the power-sensitive feature vector in the power-sensitive feature space.

[0041] After obtaining the window feature sequence, a deep nonlinear mapping transformation needs to be performed on it to extract power-sensitive features. First, the initial feature dimensions of the window feature sequence are analyzed. It is assumed that the sequence contains 128 features, and the feature values ​​follow a normal distribution with a mean of 0.5 and a standard deviation of 0.15. The numerical range of the feature values ​​is detected, and outliers exceeding the [-3, 3] interval are truncated to ensure the numerical stability of the input features.

[0042] A cascaded structure comprising four nonlinear mapping layers is constructed. The first mapping layer receives 128-dimensional input features. The feature transformation unit uses the ReLU activation function to perform an element-wise nonlinear transformation on the input. The calculation method is to compare the maximum value of each component of the input feature with zero, retaining the part greater than zero. This layer outputs a 256-dimensional intermediate feature representation. The dimensionality compression unit compresses the 256-dimensional features to 96 dimensions using a 1×1 convolutional kernel. The convolutional kernel weights are obtained through backpropagation training. The second mapping layer receives 96-dimensional features, transforms them using the ELU activation function, and outputs 192-dimensional features, which are then compressed to 64 dimensions. The third and fourth layers expand the feature dimension from 64 dimensions to 128 dimensions and then compress it to 48 dimensions, respectively, finally outputting a 32-dimensional power-sensitive feature vector.

[0043] A power consumption correlation measurement mechanism is introduced within each mapping layer. A historical power consumption fluctuation sample library is prepared, containing 10,000 sets of feature-power consumption pairing data. For the high-order feature representation output by the current layer, its cosine similarity with each set of features in the sample library is calculated. The top 50 sets of samples with the highest similarity are selected, and the power consumption fluctuation amplitudes corresponding to these samples are statistically analyzed. The Pearson correlation coefficient between each component of the current feature representation and the power consumption fluctuation amplitude is calculated. Components with an absolute correlation coefficient greater than 0.6 are marked as strongly correlated features. Strongly correlated feature components are multiplied by an enhancement factor of 1.5, and weakly correlated components with an absolute correlation coefficient less than 0.3 are multiplied by a suppression factor of 0.5. Components with correlation coefficients between the two remain unchanged in weight. After weighted adjustment, the feature vector is L2 normalized to make its magnitude 1.

[0044] During the layer-by-layer transformation process, the window feature sequence first enters the first nonlinear mapping layer. The 128-dimensional input undergoes ReLU activation and convolutional compression to generate a 96-dimensional first-layer high-order feature representation. A power correlation measurement mechanism calculates the correlation of the 96 components in this representation, identifying 38 strongly correlated components and enhancing their weights, while suppressing 24 weakly correlated components. The adjusted feature vector is then fed into the second layer, undergoing ELU activation and dimensionality transformation to generate a 64-dimensional second-layer high-order feature representation, which identifies 29 strongly correlated components. The third layer transforms the 64-dimensional features into a 48-dimensional third-layer high-order feature representation, reducing the number of strongly correlated components to 21. The fourth layer receives the 48-dimensional features and outputs a 32-dimensional final-layer high-order feature representation, retaining 18 core feature components strongly correlated with power fluctuations. The 32-dimensional vector output from the final layer is the power-sensitive feature vector mapped to the power-sensitive feature space. This vector condenses the abstract patterns most relevant to power evolution in the original window feature sequence, providing a high-quality input representation for subsequent power prediction.

[0045] In one optional implementation, the power-sensitive feature vector and the spatial feature map are fused across modalities. By constructing a feature interaction matrix to capture the spatiotemporal coupling relationship, the fused feature representation is obtained, including: The power-sensitive feature vector is decomposed in the time domain to obtain time-domain sub-feature vectors at multiple time scales, and the spatial feature map is decomposed in the spatial domain to obtain spatial sub-feature maps at multiple spatial scales. A multi-scale feature interaction matrix is ​​constructed for the temporal sub-feature vectors at each time scale and the spatial sub-feature maps at each spatial scale. A mapping relationship between the time scale and the spatial scale is established in the multi-scale feature interaction matrix, and the cross-scale interaction strength between the temporal sub-feature vectors at different time scales and the spatial sub-feature maps at different spatial scales is calculated. An adaptive attention weighting mechanism is constructed, and dynamic attention weights are assigned to the temporal sub-feature vectors of each time scale and the spatial sub-feature maps of each spatial scale according to the cross-scale interaction intensity. The dynamic attention weights are proportional to the strength of the spatiotemporal coupling relationship. Based on the dynamic attention weights, the temporal sub-feature vectors at each time scale are weighted and summed to obtain attention-enhanced temporal features, and the spatial sub-feature maps at each spatial scale are weighted and summed to obtain attention-enhanced spatial features. The attention-enhanced temporal features and the attention-enhanced spatial features are concatenated to generate a fused feature representation.

[0046] After obtaining the power-sensitive feature vector, multi-resolution wavelet decomposition is applied to the vector for time-domain decomposition. Let the dimension of the power-sensitive feature vector be d. t By setting three different decomposition levels, temporal sub-feature vectors are extracted for short-term (5-minute time window), medium-term (15-minute time window), and long-term (60-minute time window). Each temporal sub-feature vector retains the frequency component of the corresponding time scale in the original vector while filtering out interference information from other scales. Graph convolution downsampling is used to decompose the spatial feature map, dividing the original spatial feature map containing all optical module nodes into three spatial scales: local region (single rack range), medium region (single row rack range), and global region (entire data center range). Each spatial sub-feature map aggregates node features within its corresponding spatial range using graph pooling.

[0047] A multi-scale feature interaction matrix is ​​constructed, combining three temporal scales and three spatial scales to form nine spatiotemporal scale pairs. For each pair, the similarity between the corresponding temporal sub-feature vector and the spatial sub-feature map is calculated. Specifically, the temporal sub-feature vector is projected onto the same feature space as the spatial sub-feature map through a fully connected layer. The cosine similarity between the projected vector and the feature of each node in the spatial sub-feature map is calculated, and the weighted average of all node similarities is taken as the cross-scale interaction strength of that pair. The larger the value in the interaction matrix, the more significant the contribution of that spatiotemporal scale pair to power consumption prediction.

[0048] Based on the cross-scale interaction intensity values ​​in the interaction matrix, dynamic attention weights are calculated using a softmax normalization function. For the temporal domain, the attention weight for a given time scale is obtained by summing the interaction intensities across all spatial scales and then normalizing the sum. For the spatial domain, the attention weights for each spatial scale are calculated using the same method. In scenarios with sudden increases in optical module load, pairing short-term time scales with local spatial scales typically yields higher weights; in steady-state operation scenarios, pairing long-term time scales with the global spatial scale carries greater weight.

[0049] Using the calculated attention weights, a weighted sum is performed on the three temporal sub-feature vectors. Each temporal sub-feature vector is first multiplied by its corresponding attention weight, and then the three weighted vectors are summed element-wise to obtain the attention-enhanced temporal feature. This feature highlights the time-scale information most relevant to the current power consumption state. Similarly, a weighted sum is performed on the three spatial sub-feature maps to obtain the attention-enhanced spatial feature. During the weighting process, all node features in each spatial sub-feature map are uniformly multiplied by the attention weight of that spatial scale.

[0050] The attention-enhanced temporal features are flattened into a one-dimensional vector, and the node features in the attention-enhanced spatial features are aggregated into a one-dimensional vector through graph readout. These two one-dimensional vectors are then concatenated to generate a vector of dimension d. t + d s The fused feature representation encodes both the attention-filtered temporal evolution pattern and spatial topological dependency, providing comprehensive feature input for subsequent power consumption prediction. In actual deployment, the computational latency of the entire fusion process is controlled within 50 milliseconds, meeting the requirements for real-time prediction.

[0051] In one optional implementation, based on the fused feature representation and the causal dependency, the power consumption prediction value and power consumption fluctuation range are gradually derived through multi-level nonlinear transformation, including: The causal constraints of the fusion feature are generated by weighting and adjusting each feature component represented by the causal relationship according to the causal weight in the causal dependency relationship. A multi-level nonlinear extrapolation network is constructed, which includes multiple serial extrapolation layers. Each extrapolation layer performs nonlinear transformation on the input features and gradually extracts the power consumption evolution trend features. The last extrapolation layer outputs an extrapolation feature vector containing power consumption center trend information and power consumption uncertainty information. The inferred feature vector is decoupled by branching to construct a power consumption prediction branch and a fluctuation range estimation branch. The power consumption prediction branch performs nonlinear mapping on the power consumption center trend information in the inferred feature vector to obtain the power consumption prediction value. The fluctuation range estimation branch performs probability distribution modeling on the power consumption uncertainty information in the inferred feature vector to obtain the upper and lower bounds of the power consumption fluctuation range. The causal constraint fusion features are input into the multi-level nonlinear inference network for layer-by-layer nonlinear transformation, and the inference feature vector is output. The inferred feature vector is input into the power consumption prediction branch and the fluctuation range estimation branch respectively. The power consumption prediction value is obtained through the power consumption prediction branch, and the power consumption fluctuation range is obtained through the fluctuation range estimation branch.

[0052] In this technical solution, for the generation of causal constraint fusion features, a causal weight vector is first extracted from the causal dependencies. Each element of the causal weight vector corresponds to a feature component in the fusion feature representation, characterizing the degree of causal influence of that feature component on future power consumption evolution. Element-wise multiplication is performed on each feature component of the fusion feature representation, and the i-th feature component f... i With the corresponding causal weight w i Multiplying them together yields the weighted eigencomponents f. i ' = w i · f i All weighted feature components are recombined to form a causal constraint fusion feature vector. This weighting adjustment mechanism can strengthen feature components that have a significant causal effect on power consumption evolution, while suppressing interference from irrelevant features.

[0053] The multi-level nonlinear inference network is constructed using a stacked architecture, containing 3 to 5 cascaded inference layers. Each inference layer contains a fully connected group of neurons, with the number of neurons gradually decreasing from the input layer to the output layer, typically in a decreasing configuration of 256-128-64. Each inference layer uses a ReLU activation function with a leakage coefficient of 0.01 for nonlinear transformation to avoid neuron inactivation. Batch normalization is inserted between adjacent inference layers to standardize the output features of each layer, accelerating network convergence and improving numerical stability. To prevent overfitting, dropout operations are added after the second and third inference layers, with a dropout rate of 0.3. The output dimension of the final inference layer is set to 128, and the inference feature vector output by this layer contains both central tendency and dispersion information of power consumption.

[0054] To decouple the branches of the inferred feature vector, two parallel processing branches are designed. The power prediction branch uses a two-layer fully connected network. The first layer maps the 128-dimensional inferred feature vector to a 64-dimensional intermediate representation using the ReLU activation function. The second layer maps the 64-dimensional feature vector to a single scalar output without using an activation function, directly outputting the predicted power value. The fluctuation range estimation branch uses a three-layer fully connected network. The first and second layers map to 64 and 32 dimensions respectively, both using ReLU activation. The third layer outputs two values, corresponding to the upper and lower bound offsets of the fluctuation range. To ensure the reasonableness of the range, an exponential function transformation is applied to the upper bound offset to ensure it is positive, and a negative exponential transformation is applied to the lower bound offset to ensure it is negative. The upper bound of the power fluctuation range is calculated by adding the upper bound offset to the predicted power value, and the lower bound is calculated by adding the lower bound offset to the predicted power value.

[0055] In the actual inference process, the causal constraint fusion feature is input into the first layer of a multi-level nonlinear inference network. This feature vector is then sequentially processed through nonlinear transformations, batch normalization, and dropout operations at each inference layer, with the output of each layer serving as the input to the next. After processing by all inference layers, the final inference layer outputs a 128-dimensional inference feature vector. This feature vector is simultaneously input into the power consumption prediction branch and the fluctuation range estimation branch. The power consumption prediction branch directly outputs a scalar value, the predicted power consumption value in watts, through its two internal network layers. The fluctuation range estimation branch outputs two offsets through its three internal network layers. After exponential transformation, upper and lower bound offsets are obtained. These two offsets are added to the predicted power consumption value to obtain the upper and lower bounds of the power consumption fluctuation range, thus comprehensively describing the predicted range of power consumption and its uncertainty boundary.

[0056] In one optional implementation, a power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multidimensional constraint surface. The optimal operating point is then found on the multidimensional constraint surface using a gradient-guided adaptive search method, including: Collect the power consumption status parameter set and performance index parameter set of the optical module, map the power consumption status parameter set to the power consumption dimension coordinate of the power consumption-performance joint optimization space, map the performance index parameter set to the performance dimension coordinate of the power consumption-performance joint optimization space, determine the power consumption constraint boundary in the power consumption-performance joint optimization space based on the power consumption prediction value and the power consumption fluctuation range, and determine the performance constraint boundary based on the performance requirements of the optical module; The predicted power consumption value is used as the power consumption center reference point, and the upper and lower bounds of the power consumption fluctuation range are used as the power consumption deviation range. The multidimensional constraint surface is constructed in the power consumption-performance joint optimization space by combining the power consumption constraint boundary and the performance constraint boundary. A gradient-guided adaptive search method is constructed. The gradient direction of the power consumption-performance joint objective function is calculated on the multidimensional constrained surface. The power consumption-performance joint objective function comprehensively evaluates power efficiency and performance satisfaction. The search path is determined based on the gradient direction and the step size is adjusted adaptively. When the power consumption-performance joint objective function converges, the optimal operating point is determined.

[0057] After obtaining the fused feature representation and causal dependencies, the operating state of the optical module needs to be optimized and controlled. A set of power consumption state parameters is collected for the current operating state of the optical module, including instantaneous power value, average power value, power fluctuation variance, and cumulative energy consumption. Simultaneously, a set of performance index parameters is collected, including bit error rate, signal attenuation rate, transmission delay, and link stability indicators. The parameters in the power consumption state parameter set are normalized and mapped to power consumption dimension coordinates in the power consumption-performance joint optimization space. Similarly, the parameters in the performance index parameter set are normalized and mapped to performance dimension coordinates, thus constructing a joint optimization space encompassing both power consumption and performance dimensions.

[0058] In this joint optimization space, the coordinates of the power consumption center reference point are determined based on the predicted power consumption values. The upper bound of the power consumption fluctuation range is taken as the maximum allowable power consumption deviation, and the lower bound as the minimum power consumption deviation. The power consumption deviation range is determined along the power consumption dimension using the power consumption center reference point as a benchmark. Power consumption constraint boundaries are determined by combining the upper limit of the optical module's design power consumption and heat dissipation capacity constraints. Performance constraint boundaries are determined by combining the minimum bit error rate requirement, maximum allowable latency, and link availability threshold specified in the service requirements. By performing an intersection operation between the power consumption deviation range and the power consumption constraint boundaries, and simultaneously matching the performance requirements with the performance constraint boundaries, a multi-dimensional constraint surface bounded by the power consumption boundary curves and the performance boundary curves is generated in the power-performance joint optimization space. The mathematical expression of this constraint surface is as follows: and Where P(x) represents the power consumption function and Q(x) represents the performance function.

[0059] Construct a joint power consumption-performance objective function: ; Where α and β are weighting coefficients. The objective function evaluates power efficiency through the reciprocal of the power consumption term and performance satisfaction through the performance term. The gradient direction of the objective function at each point on the constraint surface is calculated. The gradient direction indicates the direction of the fastest growth of the objective function. An initial search point is selected within the constraint surface, and an iterative search is performed along the gradient direction. An adaptive step size adjustment strategy is adopted during the search process: a larger step size is used to accelerate convergence when the gradient magnitude is large, and a smaller step size is used to improve accuracy when the gradient magnitude is small. After each iteration, it is checked whether the new search point is still within the constraint surface; if it exceeds the boundary, boundary projection correction is performed. When the change in the objective function value is less than the preset convergence threshold in multiple consecutive iterations, or the gradient magnitude decreases to near zero, the search process is considered converged, and the search point at this point is the optimal operating point. The power consumption coordinates and performance coordinates corresponding to this optimal operating point determine the target operating state of the optical module, based on which operating mode switching commands and transmission parameter adjustment commands are generated.

[0060] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0061] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0062] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent prediction and dynamic optimization of power consumption of optical modules based on deep learning, characterized in that, include: Spatiotemporal decoupling processing is performed on the multidimensional operational feature data of the optical module to separate and characterize the dynamic evolution features in the time dimension and the topological correlation features in the spatial dimension, resulting in a temporal feature sequence and a spatial feature map. The multidimensional operational feature data of the optical module includes real-time acquired physical measurements such as current, voltage, temperature, optical power, and bit error rate, with a sampling frequency of 100 times per second. The correlation strength of features at different times in the time-domain feature sequence is quantified by an attention weighting mechanism, the key time window with the highest correlation strength with power consumption changes is identified, the features in the key time window are deeply embedded and transformed to generate a power consumption sensitive feature vector, and the causal dependency between the key time window and the future power consumption evolution path is established. The power-sensitive feature vector and the spatial feature map are fused across modes, and the spatiotemporal coupling relationship is captured by constructing a feature interaction matrix to obtain the fused feature representation. Based on the fused feature representation and the causal dependency, the predicted power consumption value and power consumption fluctuation range are obtained step by step through multi-level nonlinear transformation, including: The causal constraints of the fusion feature are generated by weighting and adjusting each feature component represented by the causal relationship according to the causal weight in the causal dependency relationship. A multi-level nonlinear extrapolation network is constructed, which includes multiple serial extrapolation layers. Each extrapolation layer performs nonlinear transformation on the input features and gradually extracts the power consumption evolution trend features. The last extrapolation layer outputs an extrapolation feature vector containing power consumption center trend information and power consumption uncertainty information. The inferred feature vector is decoupled by branching to construct a power consumption prediction branch and a fluctuation range estimation branch. The power consumption prediction branch performs nonlinear mapping on the power consumption center trend information in the inferred feature vector to obtain the power consumption prediction value. The fluctuation range estimation branch performs probability distribution modeling on the power consumption uncertainty information in the inferred feature vector to obtain the upper and lower bounds of the power consumption fluctuation range. The causal constraint fusion features are input into the multi-level nonlinear inference network for layer-by-layer nonlinear transformation, and the inference feature vector is output. The inferred feature vector is input into the power consumption prediction branch and the fluctuation range estimation branch respectively. The power consumption prediction value is obtained through the power consumption prediction branch, and the power consumption fluctuation range is obtained through the fluctuation range estimation branch. A power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multi-dimensional constraint surface. The optimal operating point is found on the multi-dimensional constraint surface by a gradient-guided adaptive search method. Based on the optimal operating point, an optimized control strategy containing operating mode switching instructions and transmission parameter adjustment instructions is generated.

2. The method according to claim 1, characterized in that, Spatiotemporal decoupling processing is performed on the multidimensional operational characteristic data of the optical module, separating and representing the dynamic evolution characteristics in the temporal dimension and the topological correlation characteristics in the spatial dimension, resulting in a temporal feature sequence and a spatial feature map, including: The multidimensional operational feature data is sequentially arranged according to the time axis, and the feature change trajectory within a continuous time period is extracted to form a dynamic evolution feature in the time dimension. A time-domain direction vector is constructed based on the change gradient. The multidimensional operational feature data is used to perform topological modeling based on the physical connection relationship and signal transmission path between the internal components of the optical module, extract the spatial association pattern between the components, form the topological association features in the spatial dimension, and construct the spatial direction vector based on the transmission intensity. The time-domain direction vector and the spatial-domain direction vector are orthogonalized to obtain the time-domain basis vector and the spatial-domain basis vector, which are then combined to form a time-space separation transformation matrix. The dynamic evolution features of the time dimension are multiplied by the time-domain basis vectors to obtain the time-domain projection coefficients. The time-domain projection features are then reconstructed in the time-domain subspace based on the time-domain projection coefficients. The topological correlation features of the spatial dimension are multiplied by the spatial basis vectors to obtain the spatial projection coefficients. The spatial projection features are then reconstructed in the spatial subspace based on the spatial projection coefficients. Based on the time-domain projection features, the time-domain feature sequence is organized in chronological order. Based on the spatial projection features, the spatial feature map is constructed according to the topological connection relationship.

3. The method according to claim 1, characterized in that, Deep embedding transformation is performed on the features within the key time window to generate power-sensitive feature vectors. Simultaneously, the causal dependency between the key time window and the future power evolution path is established, including: Extract all time features within the key time window and arrange them in chronological order to form a window feature sequence; The window feature sequence is subjected to multi-layer nonlinear mapping transformation. During the transformation process, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimension layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to the power consumption sensitive feature space to obtain the power consumption sensitive feature vector. Obtain historical power consumption measurement data corresponding to the key time window, as well as power consumption measurement data in the future time period after the key time window, and construct a time-aligned power consumption evolution sequence. Calculate the covariance matrix between the power-sensitive feature vector and the future power change in the power evolution sequence. Based on the covariance matrix, identify the degree of influence of each feature component in the power-sensitive feature vector on the future power evolution path. Construct a causal dependency mapping relationship from the feature components to the future power change. Quantify the causal dependency relationship between the key time window and the future power evolution path based on the causal dependency mapping relationship.

4. The method according to claim 3, characterized in that, The window feature sequence is subjected to a multi-layer nonlinear mapping transformation. During the transformation process, abstract feature representations related to power consumption fluctuation patterns are extracted layer by layer. By compressing the feature dimension layer by layer and enhancing the power consumption correlation, the window feature sequence is mapped to a power-sensitive feature space to obtain a power-sensitive feature vector, including: Obtain the initial feature dimension and feature value distribution of the window feature sequence, and determine the input features of the window feature sequence; A multi-layered cascaded nonlinear mapping layer is constructed. In each nonlinear mapping layer, a feature transformation unit and a dimension compression unit are set. The feature transformation unit extracts high-order feature representations by performing nonlinear activation operations on the input features. The dimension compression unit reduces feature redundancy by performing dimension reduction operations on the high-order feature representations. A power consumption correlation measurement mechanism is introduced in each nonlinear mapping layer to calculate the correlation coefficient between the higher-order feature representation and the historical power consumption fluctuation pattern. Based on the correlation coefficient, the weights of each feature component in the higher-order feature representation are adjusted to enhance the numerical weights of feature components that are strongly correlated with power consumption fluctuations and suppress the numerical weights of feature components that are weakly correlated with power consumption fluctuations. The window feature sequence is sequentially input into the multi-layer cascaded nonlinear mapping layer for layer-by-layer transformation. In the first nonlinear mapping layer, the window feature sequence is transformed into a first-layer high-order feature representation. In the second nonlinear mapping layer, the first-layer high-order feature representation is transformed into a second-layer high-order feature representation. This process is continued until the last nonlinear mapping layer outputs the last-layer high-order feature representation, which serves as the power-sensitive feature vector in the power-sensitive feature space.

5. The method according to claim 1, characterized in that, The power-sensitive feature vector and the spatial feature map are fused across modalities. By constructing a feature interaction matrix to capture the spatiotemporal coupling relationship, the fused feature representation is obtained, including: The power-sensitive feature vector is decomposed in the time domain to obtain time-domain sub-feature vectors at multiple time scales, and the spatial feature map is decomposed in the spatial domain to obtain spatial sub-feature maps at multiple spatial scales. A multi-scale feature interaction matrix is ​​constructed for the temporal sub-feature vectors at each time scale and the spatial sub-feature maps at each spatial scale. A mapping relationship between the time scale and the spatial scale is established in the multi-scale feature interaction matrix, and the cross-scale interaction strength between the temporal sub-feature vectors at different time scales and the spatial sub-feature maps at different spatial scales is calculated. An adaptive attention weighting mechanism is constructed, and dynamic attention weights are assigned to the temporal sub-feature vectors of each time scale and the spatial sub-feature maps of each spatial scale according to the cross-scale interaction intensity. The dynamic attention weights are proportional to the strength of the spatiotemporal coupling relationship. Based on the dynamic attention weights, the temporal sub-feature vectors at each time scale are weighted and summed to obtain attention-enhanced temporal features, and the spatial sub-feature maps at each spatial scale are weighted and summed to obtain attention-enhanced spatial features. The attention-enhanced temporal features and the attention-enhanced spatial features are concatenated to generate a fused feature representation.

6. The method according to claim 1, characterized in that, A power consumption-performance joint optimization space is constructed, in which the predicted power consumption value and the power consumption fluctuation range are mapped to a multidimensional constraint surface. The optimal operating point is then found on this multidimensional constraint surface using a gradient-guided adaptive search method, including: Collect the power consumption status parameter set and performance index parameter set of the optical module, map the power consumption status parameter set to the power consumption dimension coordinate of the power consumption-performance joint optimization space, map the performance index parameter set to the performance dimension coordinate of the power consumption-performance joint optimization space, determine the power consumption constraint boundary in the power consumption-performance joint optimization space based on the power consumption prediction value and the power consumption fluctuation range, and determine the performance constraint boundary based on the performance requirements of the optical module; The predicted power consumption value is used as the power consumption center reference point, and the upper and lower bounds of the power consumption fluctuation range are used as the power consumption deviation range. The multidimensional constraint surface is constructed in the power consumption-performance joint optimization space by combining the power consumption constraint boundary and the performance constraint boundary. A gradient-guided adaptive search method is constructed. The gradient direction of the power consumption-performance joint objective function is calculated on the multidimensional constrained surface. The power consumption-performance joint objective function comprehensively evaluates power efficiency and performance satisfaction. The search path is determined based on the gradient direction and the step size is adjusted adaptively. When the power consumption-performance joint objective function converges, the optimal operating point is determined.

7. A deep learning-based intelligent prediction and dynamic optimization system for optical module power consumption, used to implement the method as described in any one of claims 1-6, characterized in that, include: The spatiotemporal decoupling unit is used to perform spatiotemporal decoupling processing on the multidimensional operational feature data of the optical module, separating and representing the dynamic evolution features in the time dimension and the topological correlation features in the spatial dimension to obtain the temporal feature sequence and the spatial feature map. The feature weighting unit is used to quantify the correlation strength of features at different times in the time-domain feature sequence through an attention weighting mechanism, identify the key time window with the highest correlation strength with power consumption changes, perform deep embedding transformation on the features in the key time window to generate a power consumption sensitive feature vector, and establish a causal dependency relationship between the key time window and the future power consumption evolution path. A cross-modal fusion unit is used to perform cross-modal feature fusion between the power-sensitive feature vector and the spatial feature map, and to capture the spatiotemporal coupling relationship by constructing a feature interaction matrix to obtain the fused feature representation; The power consumption extrapolation unit is used to gradually extrapolate the predicted power consumption value and the power consumption fluctuation range based on the fused feature representation and the causal dependency relationship through multi-level nonlinear transformation. An optimized control unit is used to construct a power consumption-performance joint optimization space. In the power consumption-performance joint optimization space, the predicted power consumption value and the power consumption fluctuation range are mapped to a multi-dimensional constraint surface. An optimal operating point is found on the multi-dimensional constraint surface through a gradient-guided adaptive search method. Based on the optimal operating point, an optimized control strategy containing operating mode switching instructions and transmission parameter adjustment instructions is generated.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images