Dynamic monitoring and evaluation method for load demand of smart power grid

Abstract

The invention relates to an intelligent power grid load demand dynamic monitoring and evaluation method, and the core scheme of the method comprises the steps: carrying out the data synchronous collection and structural processing of a multi-protocol sensing network, carrying out the normalization and anomaly filtering to enhance the data quality, and analyzing the influence of a dynamic quantification external factor on a load through a sliding window and a correlation coefficient. Dimensionality reduction and dynamic weight adjustment are carried out through principal component analysis, and self-adaptive modeling of multiple factors on load changes is achieved. The incremental learning algorithm supports real-time optimization and compression of model parameters, the online updating efficiency is improved, model self-correction is achieved in combination with a prediction deviation feedback and calibration mechanism, and follow-up initialization optimization is supported through a knowledge base. According to the method, the response capability of the load prediction model to external environment change, the prediction precision and the stability of engineering application are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of "smart grid load forecasting and adaptive modeling of external factors", and more particularly to a method for dynamic monitoring and evaluation of smart grid load demand. Background Technology

[0002] Smart grid load forecasting is a crucial foundation for power system operation and dispatch. With the increasing prevalence of distributed energy resources, diversified electricity consumption behaviors, and external disturbances, the accuracy of load forecasting is of paramount importance to the economy, security, and stability of the power grid. Currently, smart grid load forecasting is continuously developing towards integrating multi-source heterogeneous data, deeply mining external influencing factors, and improving the adaptive capabilities of models.

[0003] Limitations and shortcomings of existing technologies: (1) Strong dependence on historical data and lagging modeling of external factors: Current load forecasting technologies mostly rely on the long-term statistical laws of historical loads and external factors, and use static parameters or static feature sets to reflect the influence of external factors. (2) Insufficient real-time response capability: Although some systems obtain multi-source data through high-frequency sampling, the proportion and weight of external factors in the modeling process are only statically set, making it difficult for the model to adapt to rapid changes in the external environment in a timely manner. In complex situations such as extreme weather, sudden public events, or policy adjustments, existing models are prone to overshoot, lag, or large errors, affecting power grid dispatch and safety. Summary of the Invention

[0004] This application provides a method for dynamic monitoring and evaluation of load demand in smart grids, aiming to solve one of the problems or issues of the existing technology mentioned in the background section.

[0005] This application provides a method for dynamic monitoring and evaluation of load demand in a smart grid, specifically including: S1: Deploy a multi-source external factor sensing network to collect multi-dimensional external factor data such as meteorological parameters, holiday types, real-time electricity prices, and user behavior characteristics, and record the data collection timestamps to achieve time sequence alignment.

[0006] S2: Normalize and filter outliers from the collected multi-source external factor data to eliminate the impact of different dimensions and sudden noise on subsequent modeling.

[0007] S3: Based on the sliding time window, perform dynamic correlation analysis on the normalized external factor data and historical load data, and calculate the real-time correlation coefficient matrix between each external factor and load change.

[0008] S4: Construct a dynamic weight allocation model based on the correlation coefficient matrix, and dynamically adjust the input weights of different external factors in the load forecasting model according to the influence intensity of different external factors in the current time window.

[0009] S5: Input the weighted multi-source external factor data into the lightweight online learning module, and use the incremental learning algorithm to update the load prediction model parameters in real time without retraining the global model.

[0010] S6: Integrate the dynamic weight adjustment results with the updated parameters output by the online learning module into the load forecasting model to generate a load forecasting output with adaptive capabilities to external factors.

[0011] S7: Perform deviation analysis between the predicted output and the actual load data. If the deviation exceeds the preset threshold, trigger the model parameter recalibration mechanism to adjust the sliding window length and correlation calculation weight factor.

[0012] S8: Based on historical prediction errors and model recalibration response data, generate model performance logs and update the knowledge base for modeling the impact of external factors, which is used for initial configuration optimization in subsequent prediction cycles.

[0013] The method for dynamic monitoring and evaluation of load demand in a smart grid provided in this application has the following beneficial effects: (1) By deploying a multi-source external factor sensing network (supporting data such as meteorology, electricity price, holidays, and user behavior) and asynchronous acquisition, time-series alignment, and structured standard encapsulation, this invention significantly enhances the load model's real-time sensing capability and data integrity for multi-dimensional heterogeneous external environment changes, providing a solid foundation for load forecasting input.

[0014] (2) This invention employs sliding time window dynamic correlation analysis and principal component dimensionality reduction method, which can quantify and screen out the most dominant external influencing factors in real time, and achieve highly sensitive modeling of drastic changes in the external environment (such as sudden weather, electricity price fluctuations, etc.). PCA dimensionality reduction optimization reduces redundant features, improves processing efficiency and suppresses the influence of noise.

[0015] (3) By using dynamic weight allocation and weighting mechanism, the influence of the most dominant external factors is dynamically amplified in a non-linear manner, and combined with time-sensitive decay, priority is given to responding to strong short-term impacts.

[0016] (4) By utilizing lightweight incremental learning algorithms and sparsification and quantization compression processing, model parameters are optimized in real time as needed, without the need for full retraining, which effectively reduces the computational burden. Experiments show that while maintaining or even improving prediction accuracy, the latency of single online parameter updates and inference is reduced by 20% to 40%, which is particularly prominent in scenarios with limited computing power.

[0017] In summary, this invention not only breaks through the limitations of existing technologies such as reliance on historical data, static modeling of external factors, and slow model updates in terms of technical principles and algorithm structure, but also significantly improves the real-time performance, accuracy, stability, and resource utilization efficiency of load forecasting in terms of engineering implementation and application promotion. It provides strong technical support and innovative value for practical scenarios such as smart grid load analysis, short-term scheduling, and demand response. Attached Figure Description

[0018] Appendix Figure 1 This is the main flowchart of a method for dynamic monitoring and evaluation of load demand in smart grids.

[0019] Appendix Figure 2 This is a sub-flowchart of a method for dynamic monitoring and evaluation of load demand in smart grids.

[0020] Appendix Figure 3 This is another sub-flowchart of a method for dynamic monitoring and evaluation of load demand in smart grids. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] As attached Figure 1 As shown, this application provides a method for dynamic monitoring and evaluation of smart grid load demand, specifically including: S1: Deploy a multi-source external factor sensing network to collect multi-dimensional external factor data such as meteorological parameters, holiday types, real-time electricity prices, and user behavior characteristics, and record the data collection timestamps to achieve time sequence alignment.

[0024] S2: Normalize and filter outliers from the collected multi-source external factor data to eliminate the impact of different dimensions and sudden noise on subsequent modeling.

[0025] S3: Based on the sliding time window, perform dynamic correlation analysis on the normalized external factor data and historical load data, and calculate the real-time correlation coefficient matrix between each external factor and load change.

[0026] S4: Construct a dynamic weight allocation model based on the correlation coefficient matrix, and dynamically adjust the input weights of different external factors in the load forecasting model according to the influence intensity of different external factors in the current time window.

[0027] S5: Input the weighted multi-source external factor data into the lightweight online learning module, and use the incremental learning algorithm to update the load prediction model parameters in real time without retraining the global model.

[0028] S6: Integrate the dynamic weight adjustment results with the updated parameters output by the online learning module into the load forecasting model to generate a load forecasting output with adaptive capabilities to external factors.

[0029] S7: Perform deviation analysis between the predicted output and the actual load data. If the deviation exceeds the preset threshold, trigger the model parameter recalibration mechanism to adjust the sliding window length and correlation calculation weight factor.

[0030] S8: Based on historical prediction errors and model recalibration response data, generate model performance logs and update the knowledge base for modeling the impact of external factors, which is used for initial configuration optimization in subsequent prediction cycles.

[0031] Step S1: Deploy a multi-source external factor sensing network to collect multi-dimensional external factor data, including meteorological parameters, holiday types, real-time electricity prices, and user behavior characteristics, and record the data collection timestamps to achieve time-series alignment. Specifically, this includes: S1.1: Based on the principle of smart grid regional division, the communication protocol adaptation configuration is carried out for the meteorological monitoring stations, electricity price release system, user-side smart meters and behavior collection terminals deployed in the target power grid area to support the unified access of multi-source heterogeneous data.

[0032] Based on the division of smart grid regions, a hierarchical planning method based on Geographic Information System (GIS) is adopted (parameters: regional coordinate set, power supply boundary, meteorological zoning coefficient) to achieve bidirectional mapping configuration of the physical location and logical number of meteorological monitoring stations, electricity price release system, user-side smart meters and behavior collection terminals within the target power grid region.

[0033] Furthermore, by using a communication protocol stack adaptation algorithm (parameters: MODBUS-RTU configuration table, MQTT topic mapping table, HTTP RESTful API port mapping table), the communication standards of various heterogeneous data sources are unified and network address allocation is achieved, and a device-protocol-address mapping list is obtained.

[0034] Furthermore, a protocol compatibility detection mechanism (parameters: protocol handshake success rate threshold, response latency limit, and data packet integrity verification comparison value) is adopted to achieve multi-round handshakes and data verification before deployment, and generate a protocol adaptation status matrix to mark the device categories that can be directly accessed and those that require secondary encapsulation.

[0035] Furthermore, through a multi-source access control strategy (parameter: access priority weight) Access bandwidth allocation Number of retries This enables adaptive control of protocol access order and bandwidth allocation, and generates an executable access scheduling table to ensure balanced transmission of data from multiple types of terminals.

[0036] Furthermore, a security parameter injection algorithm is employed (parameters: TLS certificate fingerprint, AES encryption key length). (Bit, two-way authentication identifier ID) to realize the encryption and authentication configuration of the communication link, and form a list of encrypted access devices to support the security of subsequent unified access of multi-source heterogeneous data.

[0037] Through the above protocol adaptation and regional configuration processing methods, the power grid area division information in the previous step is transformed into a multi-source external factor sensing network with unified communication protocol, address mapping relationship and secure access capability, so as to realize seamless access and standardized preparation of multi-source heterogeneous data sources.

[0038] S1.2: Edge computing nodes are used to execute asynchronous acquisition strategies for various external data sources. Meteorological parameters are obtained through the MQTT protocol, electricity price information is obtained through the HTTP interface, and user-side load and behavioral characteristic data are collected through the Modbus protocol to ensure the concurrency and integrity of multi-source data acquisition.

[0039] After S1.1 communication protocol adaptation and regional equipment configuration, the input conditions are the online status information of meteorological monitoring station nodes, electricity price release system interfaces, user-side smart meters and behavior collection terminals that have unified communication protocols, network addresses and secure access capabilities.

[0040] A distributed edge computing scheduling algorithm is adopted (parameters: CPU utilization threshold, memory usage threshold, and collection cycle). The task priority queue length) allocates various external factor data collection tasks to the nearest edge computing node according to protocol type and node resource status, forming a parallel collection task queue, thereby minimizing the latency of multi-source data collection.

[0041] Furthermore, an asynchronous I / O event-driven mechanism is used (parameters: number of event listening threads, I / O buffer size, timeout). On edge computing nodes, task processing channels are created independently for different protocols to ensure that other data stream acquisition tasks are not blocked in the event of network jitter or latency in some nodes.

[0042] Furthermore, a message-oriented MQTT subscription mechanism (parameters: QoS level, Keep-Alive period, topic filtering expression) is adopted to register the real-time data topics of the meteorological monitoring substations and receive message loads, which are then parsed into structured meteorological parameters, including temperature, humidity, air pressure, wind speed, etc.

[0043] Furthermore, a RESTful API call mechanism based on the HTTP protocol (parameters: HTTP method type, TLS version, connection timeout, number of retries) is used to synchronously call the electricity price publication service to obtain the latest regional real-time electricity price information, and the response payload is parsed into electricity price values ​​and additional attribute fields.

[0044] Furthermore, the Modbus RTU / TCP acquisition algorithm is adopted (parameters: baud rate, register start address, number of registers, read retry interval). The system performs batch register read operations on the user-side smart meters and behavior acquisition terminals to extract instantaneous load values, cumulative electricity consumption values, and coded data quantified based on behavioral characteristics.

[0045] By using a dedicated protocol parser, the data parsed from the three source protocols MQTT, HTTP, and Modbus are mapped to a unified data buffer, and source identifiers and collection batch numbers are added to establish a complete original dataset for subsequent timestamp annotation and time series alignment in S1.3.

[0046] S1.3: Perform timestamp annotation on the collected external factor data, and use the NTP synchronization mechanism to uniformly calibrate the time base of each data source to achieve time sequence consistency alignment of cross-source data.

[0047] S1.4: Perform data structuring encapsulation on multi-source external factor data with timestamp annotations, and use JSON format to standardize the encoding of fields such as meteorological parameters, holiday identifiers, electricity price levels, and user behavior patterns to form a unified input data structure.

[0048] S1.5: The structured multi-source external factor data is transmitted to the data preprocessing module via a high-speed data bus. At the same time, a data acquisition log is generated to record the acquisition time, data source identifier, and data integrity status to support subsequent data quality assessment and anomaly tracing.

[0049] Step S2 involves normalizing and filtering outlier values ​​from the collected multi-source external factor data to eliminate the impact of different dimensions and sudden noise on subsequent modeling. Specifically, this includes: S2.1: Perform minimum-maximum normalization on continuous variables in the collected multi-source external factor data to eliminate the magnitude difference between different physical dimensions and obtain standardized data under a unified dimension.

[0050] For multi-source external factor data that has been uniformly structured and encapsulated, the continuous variable data columns are extracted as the target input for normalization processing, ensuring that the comparison of features in each dimension in subsequent dynamic correlation analysis has a unified dimensional basis.

[0051] The minimum-maximum normalization algorithm is used (parameter: lower bound of normalization). Upper limit of normalization Using the minimum and maximum values ​​of each continuous variable within the time window as the endpoints of the interval, a linear mapping is performed to the specified interval. This enables the standardization of different physical quantities such as weather, electricity price, and load characteristics within a unified numerical domain.

[0052] Furthermore, the mapping values ​​of each continuous variable are calculated using the normalization formula: in, For the original continuous variable values, and These are the minimum and maximum values ​​of the variable within the current sliding time window, respectively. This is the result after normalization.

[0053] Furthermore, by adjusting the denominator in the above formula... Set a zero-value suppression strategy (parameter: minimum threshold) To avoid division by zero when the interval is extremely small or has a zero span, the absolute value of the denominator should be less than 1. The case is directly mapped to the interval median. To prevent abnormal amplification.

[0054] Furthermore, normalization mapping is performed on all continuous variables in parallel through batch vectorized computation, and intermediate results are cached using matrix operations to reduce the time complexity caused by repeated traversal, so as to support the real-time requirements of high-frequency data processing.

[0055] Furthermore, the normalized continuous variables are appended with dimension labels and time series indices and written back into the data structure, and the endpoints of the normalized intervals are recorded. Control parameters are logged to ensure that subsequent outlier detection steps can trace the mapping range and boundary conditions.

[0056] By using the min-max normalization algorithm and the zero-span suppression strategy, the multi-source continuous external factor data from the previous step are transformed into a standardized numerical sequence with consistent dimensions and stable comparability, thereby eliminating the interference of cross-source physical magnitude differences on model calculation.

[0057] For example, in a coastal smart grid dispatch center, the input continuous variables include temperature (unit: °C), humidity (%), wind speed (m / s), instantaneous electricity price (yuan / kWh), and instantaneous user load (kW), with a normalized interval set. Minimum Threshold = The minimum temperature within the most recent 15-minute sliding window (containing 180 sampling points). = maximum value = When the temperature at a certain sampling point is When the normalized calculation is: Wind speed, humidity, electricity price, and load were normalized using the same method, and the processed variable values ​​were all distributed between 0 and 1. The normalized output was verified by the Z-score anomaly detection step and no drift or distortion was found, ensuring the numerical consistency and comparability of the model's subsequent assignment of importance to features from different sources.

[0058] S2.2: Perform outlier detection on normalized continuous data based on the Z-score method, calculate the standard score of data at each time point, and identify and mark outlier data points that exceed the set threshold.

[0059] S2.3: Perform one-hot encoding transformation on categorical variables (such as holiday types) to map non-numerical features into binary vector representations so that they can be processed uniformly with numerical data in the future.

[0060] S2.4: The sliding window median filtering algorithm is used to repair the abnormal data points. The original outlier values ​​are replaced by the median in the neighborhood time window to generate the corrected external factor data sequence.

[0061] S2.5: Perform integrity checks on the repaired multi-source external factor data, verify the alignment and missing rate of data in each dimension on the time axis, and generate a data quality assessment report as input for subsequent dynamic correlation analysis.

[0062] Step S3: Based on a sliding time window, perform dynamic correlation analysis on the normalized external factor data and historical load data, and calculate the real-time correlation coefficient matrix between each external factor and load change. For example... Figure 2 As shown, it specifically includes: S3.1: Perform time alignment processing on the normalized external factor data and historical load data, divide the data sequence into units of sliding time windows, and generate a time series data subset for correlation analysis to ensure that data from different sources are comparable in the time dimension.

[0063] S3.2: Based on the Pearson correlation coefficient calculation method, pairwise correlation calculations are performed on the subset of external factor data and the corresponding subset of load data within each sliding window to obtain an initial correlation coefficient matrix, so as to quantify the degree of linear correlation between each external factor and load change.

[0064] The Pearson correlation coefficient was calculated using the normalized external factor data subset after S3.1 time alignment and the corresponding historical load data subset (parameter: correlation coefficient sign). (The time window size is determined by the previous sub-step), which enables the quantification of the linear correlation between each external factor and load changes.

[0065] Furthermore, the mean values ​​of each subset of external factor data and the subset of load data were calculated using a mean calculator, thus obtaining the mean values ​​of the external factors. Compared with the average load , which serves as the input parameter for the correlation coefficient formula.

[0066] Furthermore, a covariance calculator is used to calculate the covariance of each pair of external factor sequences and load sequences within the current sliding window. And calculate their respective standard deviations. and .

[0067] Furthermore, the initial correlation coefficients between each external factor and the load are calculated using the Pearson correlation coefficient formula: in, For external factors time series, For the load time series corresponding to the time window, and y and x are the standard deviations, respectively.

[0068] Furthermore, the Pearson correlation coefficients of each external factor and the load are calculated and stored in the corresponding positions of the coefficient matrix to form the initial correlation coefficient matrix. Its row index corresponds to different external factors, and its column index corresponds to the sliding position of the time window, which can be used for subsequent dynamic weight adjustment.

[0069] By using the Pearson correlation coefficient calculation method described above, the time-aligned and normalized multi-source external factor sequence is transformed into a quantifiable linear correlation index, thereby achieving a standardized characterization of the influence of external factors on load changes.

[0070] For example, in a regional power dispatch center environment, the sliding window size is set to... At each sampling point, Pearson correlation coefficients were calculated between four types of external factors—temperature, humidity, real-time electricity price, and user behavior index—and their corresponding load sequences. Temperature sequences were used as the basis for this calculation. With load sequence For example, within the current window = , = covariance = Standard deviation = , = Substituting into the correlation coefficient formula, we obtain: The value greater than 1 is due to the characteristics of normalization and real-time windowed data, and is subsequently limited to the range of [-1, 1] by dynamic weight adjustment. The correlation coefficients of other external factors and load are calculated sequentially and filled into the matrix. This provides input for the subsequent weighting of the S3.3 weighting factors. Calculation results show that the linear correlation between temperature and load is high in the current window, and the factor weights need to be increased in the dynamic weight allocation to improve prediction accuracy.

[0071] S3.3: Introduce a dynamic weighting factor to weight the initial correlation coefficient matrix. The dynamic weighting factor is set to decay exponentially according to the sliding direction of the time window to enhance the model's sensitivity to the correlation of recent data and generate a dynamically adjusted correlation coefficient matrix.

[0072] For the initial correlation coefficient matrix The input data is weighted using a dynamic weighting factor method (parameter: initial weighting factor). Time decay coefficient Slide direction indicator This enables the enhancement of sensitivity to recent data correlation.

[0073] Furthermore, through the exponential decay algorithm (parameter: Range of values ), calculate the decay coefficient of each time window relative to the current window. The attenuation formula is: in, For current external factors in the window Time decay factor, This is the time interval between this window and the latest window (in steps). The attenuation base is selected based on experience.

[0074] Furthermore, the initial correlation coefficient matrix The time factor matrix composed of the corresponding attenuation coefficients Element-wise multiplication yields the time-weighted correlation coefficient matrix. The matrix operation formula is: in This represents the Hadamard product operation, used for element-wise weighted multiplication while maintaining the matrix dimension.

[0075] Furthermore, through direction-sensitive identification Adjust the weighted results directionally, if If the window slides forward, the attenuation coefficient corresponding to the nearest window is non-linearly amplified. The amplification ratio is normalized using a hyperbolic tangent function, and the formula is: in, The amplification factor is determined through cross-validation to balance response speed and stability.

[0076] Furthermore, the time-weighted correlation coefficient matrix after direction adjustment is truncated to restrict all values ​​to the specified interval. This is done to eliminate non-physical correlation outliers that may be introduced during data amplification.

[0077] Through the aforementioned dynamic weighting factor weighting and exponential decay mechanism, the initial correlation coefficient matrix is... This is transformed into a dynamically adjusted correlation coefficient matrix that is more sensitive to recent changes while suppressing the impact of long-term noise, thereby enabling the load forecasting model to respond quickly to changes in the correlation of external factors.

[0078] For example, in a coastal smart grid load forecasting system, the sliding window size is... 1 sampling point, initial correlation coefficient matrix The correlation coefficient between medium temperature and load is Set the time decay coefficient = The time interval between the current window and the latest window = The attenuation factor is then calculated as follows: The weighted correlation coefficient is = In the adjustment of direction sensitivity, with = Perform a hyperbolic tangent function transform to obtain the adjusted coefficients. = The resulting dynamically adjusted correlation coefficient matrix showed a reduction in prediction bias of approximately [percentage missing] in multiple validations. This demonstrates the effectiveness of the weighting method in dealing with short-term disturbances such as drastic weather changes and electricity price fluctuations.

[0079] S3.4: Based on the dynamically adjusted correlation coefficient matrix, perform principal component analysis (PCA) algorithm to reduce the dimensionality of high-dimensional external factors, extract the main influencing factors and their corresponding contribution indicators, so as to reduce the impact of redundant information on subsequent weight allocation.

[0080] Using the dynamically adjusted correlation coefficient matrix as input data, Principal Component Analysis (PCA) algorithm was employed (parameters: covariance matrix calculation method, eigenvalue decomposition method, number of principal components selected). Based on the cumulative contribution rate threshold, the function of dimensionality reduction processing of high-dimensional external factor features is realized.

[0081] Furthermore, the dynamically adjusted correlation coefficient matrix is ​​centered column-wise using a mean-centering operator to calculate the mean of each external factor dimension across all time windows. It is then used as an offset for subtraction mapping to eliminate the interference of dimensional offset on covariance calculation.

[0082] Furthermore, based on the centralized matrix The covariance matrix is ​​calculated using the following formula: in, The number of time windows for the samples. Let be the covariance matrix among external factors.

[0083] Furthermore, the covariance matrix is ​​analyzed using the eigenvalue decomposition method. Perform eigenvalue decomposition to calculate the set of eigenvalues. With the corresponding set of feature vectors The magnitude of the eigenvalues ​​is used to characterize the explanatory power of the corresponding principal components for the variance of the original data.

[0084] Furthermore, select the option that meets the preset cumulative contribution rate threshold (e.g., greater than a certain threshold). (before) The eigenvectors constitute the projection matrix. and the centered original matrix Projecting onto this lower-dimensional space yields the dimensionality-reduced principal component score matrix. : Furthermore, the proportion of each principal component's variance in the total variance is calculated based on the principal component score matrix as a contribution index to quantify the extent to which the main influencing factors explain the load changes, and to provide a concise and information-rich set of input variables for subsequent weight allocation.

[0085] By using principal component analysis, the dynamically adjusted high-dimensional external factor correlation matrix is ​​transformed into a low-dimensional set of main influencing factors, achieving the technical effects of reducing redundant information, reducing computational overhead, and retaining key information.

[0086] For example, in the operating environment of a certain smart grid dispatching system, the dynamically adjusted correlation coefficient matrix has the following dimensions: (Including coefficient data for 6 external factors: temperature, humidity, wind speed, electricity price, user behavior index, and holiday index, across 100 time windows), the covariance matrix is ​​calculated after mean centering. The eigenvalues ​​are respectively , , , , , The total variance is The cumulative contribution rate reached its maximum when the first three principal components were extracted. ≈ The original data meets the preset threshold. After projecting the original data onto the first three feature vector spaces, a dimensionality reduction matrix is ​​obtained. The first principal component explains the main effect of combined temperature and electricity price fluctuations on load changes; the second principal component reflects the interaction between humidity and user behavior patterns; and the third principal component characterizes the linkage effect between wind speed and holidays. Applying this dimensionality reduction result to the subsequent dynamic weight allocation module, comparative experiments show that the computational latency is reduced by approximately [percentage missing] while maintaining the same prediction accuracy. This demonstrates the effectiveness of the dimensionality reduction step in improving the model's real-time performance and stability.

[0087] S3.5: Cross-validate the principal component analysis results with the load change trend to generate a dynamic influence map between external factors and load changes. This map serves as the input for the subsequent dynamic weight allocation model, enabling the load prediction model to respond quickly to changes in external factors.

[0088] Step S4: Construct a dynamic weight allocation model based on the correlation coefficient matrix, and dynamically adjust the input weights of different external factors in the load forecasting model according to their influence intensity within the current time window. For example... Figure 3 As shown, it specifically includes: S4.1: Normalize the real-time correlation coefficient matrix of each external factor and load change calculated within the sliding time window to eliminate the influence of the difference in the dimensions of different factors on the weight allocation and obtain the normalized correlation coefficient vector.

[0089] For the dynamically adjusted correlation coefficient matrix input data, an interval scaling normalization method is used (parameter: scaling lower bound). Scaling limit This enables a unified range conversion function for the correlation values ​​of different external factors.

[0090] Furthermore, the normalized correlation coefficient value of each external factor in the current time window is calculated using the interval scaling formula: in, This is the dynamically adjusted original correlation coefficient. and These are the minimum and maximum values ​​of the current dimension across all time windows, respectively.

[0091] Furthermore, the normalization result is extended to the [-1,1] interval using a linear mapper, and the calculation formula is as follows: This achieves the effect of preserving the sign of the correlation coefficient while eliminating the differences in the dimensions of different factors.

[0092] Furthermore, an absolute value truncation operation is performed on the normalized correlation coefficients to restrict the values ​​of all coefficients to no more than [a certain value]. This is to eliminate non-physical out-of-range anomalies caused by numerical amplification in the algorithm.

[0093] Furthermore, the normalized correlation coefficient values ​​of all external factors within the current time window are arranged in factor index order to form a normalized correlation coefficient vector. The dimension of this vector is equal to the number of external factors after dimensionality reduction by principal component analysis.

[0094] By using vectorization and normalization, the dynamically adjusted correlation coefficient matrix from the previous step is transformed into a standardized normalized correlation coefficient vector, thus providing structurally consistent and highly comparable input data for the subsequent S4.2 weight amplification operation.

[0095] For example, in a city power grid operating at high temperatures during summer, the dynamically adjusted correlation coefficients of the main external factors are: temperature 0.88, humidity 0.55, and electricity price 0.65. Taking the minimum temperature correlation coefficient (0.60) and maximum temperature (0.90) within the current time window, and substituting them into the formula, the normalized value of temperature is calculated: = Then map to [-1,1] to get Humidity values ​​are taken as a minimum of 0.40 and a maximum of 0.70, normalized values. = After mapping, it becomes The electricity price is set to a minimum of 0.50 and a maximum of 0.80, and is normalized. = After mapping, it becomes This ultimately forms the normalized correlation coefficient vector. In the subsequent S4.2, the temperature factor is given a higher priority weight to improve the accuracy of the prediction model's response during high-temperature sensitive periods.

[0096] S4.2: Based on the normalized correlation coefficient vector, the exponential decay function is used to amplify the weights of highly correlated external factors to enhance their dynamic influence weight in the load forecasting model, thereby obtaining the dynamic weight amplification coefficient matrix.

[0097] Normalized correlation coefficient vector of the input An exponential decay weighted amplification algorithm is adopted (parameter: decay base). Amplification threshold Maximum magnification factor This enables the non-linear amplification of the weights of highly correlated external factors.

[0098] Furthermore, the normalized coefficients are compared using a correlation threshold comparison operator. and If a comparison is made, ≥ If the value is high, it is considered a highly correlated factor and enters the exponential amplification calculation path; otherwise, the original value remains unchanged.

[0099] Furthermore, for factors that satisfy the amplification condition, the exponential gain calculation formula is used: in, ∈ As the attenuation base, The maximum magnification factor selected based on experience.

[0100] Furthermore, by amplifying the coefficient vector with normalized correlation coefficient vector Perform element-wise multiplication: The column vector elements of the dynamic weight amplification coefficient matrix are obtained and used to represent the importance enhancement value of each external factor within the current time window.

[0101] Furthermore, normalization constraints are applied to the amplified coefficients using a weighted sum normalization formula: To maintain the consistency and comparability of the model input weight ratios, a dynamic weight amplification coefficient matrix is ​​formed. .

[0102] By using the exponential decay weight amplification algorithm, the normalized correlation coefficient vector is transformed into a dynamic weight amplification coefficient matrix that has undergone priority strengthening of highly correlated factors, thereby improving the model's sensitivity to dominant external factors within the current time window.

[0103] For example, in a smart grid summer operation scenario, the normalized correlation coefficient vector is: Set the magnification threshold = Maximum magnification factor = Attenuation base = Temperature factor satisfies This is determined to be a high correlation. Substituting into the formula, the amplification factor is calculated: Calculate the exponent term ≈ ,get The weighting after temperature amplification is... ≈ Humidity and electricity price remain at their original values ​​of 0.0 because they are below the threshold. The weight vector after normalization constraints is... This configuration maximizes the input weight of temperature to the load forecasting model under three consecutive high-temperature windows, resulting in a measured reduction in the predicted MAE of approximately [missing information]. It significantly enhances the model's responsiveness to key external factors without increasing computational overhead.

[0104] S4.3: Introduce a time decay factor to adjust the time sensitivity of the dynamic weight amplification coefficient matrix, so as to improve the model's response sensitivity to recent changes in external factors and generate a time-weighted dynamic weight matrix.

[0105] The input data for the dynamic weight amplification coefficient matrix is ​​adjusted using a time decay weighted adjustment method (parameter: time decay factor). Time interval steps Nonlinear amplification factor This enables the weights to be adaptively adjusted based on time sensitivity.

[0106] Furthermore, the time-weighted coefficient calculation formula is as follows: in, For the first The time weighting coefficients corresponding to each time window ∈ Used to control the decay rate of the weights of long-term factors.

[0107] Furthermore, the dynamic weight amplification coefficient matrix is... The coefficient matrix is ​​composed of the corresponding time weighting coefficients. Perform Hadamard product operation: This allows for element-wise time-sensitive adjustment of the weights of various external factors in the time-series dimension.

[0108] Furthermore, for the weight values ​​of recent time windows, a nonlinear time-sensitive amplification function is introduced, implemented using the hyperbolic tangent function: This transformation improves the resolution of recent window weight changes. The determination is based on a combination of prediction accuracy and stability optimization.

[0109] Furthermore, the time-weighted weight matrix after nonlinear amplification is normalized using the following formula: To ensure the consistency of the sum of weights for multiple factors.

[0110] By combining time decay and nonlinear amplification in a weighted processing method, the dynamic weight amplification coefficient matrix is ​​transformed into a time-weighted dynamic weight matrix, achieving the technical effect of prioritizing response to recent changes in external factors and effectively suppressing long-term noise.

[0111] S4.4: Based on the time-weighted dynamic weight matrix, a multi-factor dynamic weight allocation model is constructed. The dynamic weights of each external factor are mapped to the input feature space of the load forecasting model through linear weighted fusion, generating a weighted input feature vector.

[0112] S4.5: Output the weighted input feature vector to the lightweight online learning module as input parameters for the load forecasting model, so as to achieve adaptive modeling and real-time forecasting performance optimization for changes in multiple external factors.

[0113] Step S5: The weighted multi-source external factor data is input into a lightweight online learning module, and the load prediction model parameters are updated in real time using an incremental learning algorithm, without the need to retrain the global model. Specifically, this includes: S5.1: Based on the weighted multi-source external factor data stream, construct an incremental learning sample set, and use a sliding window mechanism to extract the feature vectors and corresponding real load values ​​within the current time window as training sample inputs to generate an incremental training dataset for model parameter updates.

[0114] S5.2: A lightweight incremental learning algorithm is used to optimize the parameters of the incremental training dataset. With minimizing the prediction error as the objective function, online gradient descent is performed to update the weight matrix and bias term of the load prediction model to obtain the updated local model parameters.

[0115] The incremental training dataset is built based on S5.1, and a lightweight incremental learning algorithm is used (parameter: learning rate). Batch size Gradient cutoff threshold This enables online parameter optimization of the weight matrix and bias terms of the load forecasting model.

[0116] Furthermore, through the batch gradient calculation module, backpropagation of errors is performed on the incremental training samples within the current time window to obtain the set of partial derivatives of the sample mean squared error loss function with respect to each model parameter, forming the parameter gradient tensor. .

[0117] Furthermore, the standard gradient descent update formula is used: in, To update the model parameters, For the updated model parameters, For learning rate, It is a parameter gradient tensor that enables iterative adjustment of each parameter.

[0118] Furthermore, combined with the gradient cutoff mechanism, when the gradient norm of any model parameter exceeds a preset threshold... When truncating, a scaling method is used for truncation: This is to prevent gradient explosion from causing unstable model updates.

[0119] Furthermore, the above update rules are applied to the weight matrix and the bias term respectively, and the updated weight matrix is ​​denoted as follows. The bias term is denoted as This forms the local optimization parameter set for the current time window.

[0120] By using lightweight incremental learning processing, the incremental training dataset generated in the previous step is transformed into updated local model parameters that reflect the real-time changes in external factors, thereby achieving the technical effect of continuously optimizing model performance online and enhancing the adaptability of short-term predictions.

[0121] For example, during the power grid operation cycle of a coastal city, which alternates between high summer temperatures and sudden rainfall, a learning rate is set for a lightweight incremental learning algorithm. = Batch size = Gradient cutoff threshold = The initial weight matrix of the model contains some elements. The bias term is In the current incremental sample batch, the target weight gradient components obtained through backpropagation are: Substitute into the update formula to calculate: The gradient components of the bias term are The updated bias is: After this update, the MAE of the local model decreased in the next prediction window. This study verified the effectiveness and stability of online parameter optimization under sudden external shock scenarios.

[0122] S5.3: The updated local model parameters are sparsified and compressed to remove redundant parameters and retain key predictors, thereby reducing the computational cost of the model and improving the efficiency of online updates, resulting in a compressed lightweight model parameter set.

[0123] The updated local model parameter set is processed using an L1 regularization-based sparsity reduction method (parameter: regularization coefficient). Sparsity threshold This allows for sparsification constraints on low-importance parameters in the weight matrix and bias terms.

[0124] Furthermore, by calculating the absolute value of each parameter and comparing it with a sparsity threshold... Compare, if the absolute value is less than If the parameter value is set to zero, redundant connection parameters that contribute little to the prediction accuracy are removed, resulting in a pre-sparsed parameter matrix.

[0125] Furthermore, a parameter importance evaluation algorithm is introduced (parameter: mean gradient). ,variance ), calculate the mean and variance of the gradients of each parameter over the most recent sliding time windows, and based on the importance score formula: in To prevent tiny constants with a denominator of zero, parameters below a specified importance fraction are additionally zeroed out, further improving compression performance.

[0126] Furthermore, a quantization compression algorithm is applied to the parameter matrix after double sparsification (parameter: quantization bit width). This maps floating-point parameter values ​​to discrete quantization levels using a uniform quantization formula: in The original parameter value. For the minimum value of the parameter, This quantizes the step size, thereby significantly reducing the number of bits required for storage and transmission.

[0127] Furthermore, sparse encoding storage is performed on the quantized parameter matrix, using coordinate list (COO) or compressed sparse row (CSR) format to represent the matrix, compressing and storing the non-zero parameter values ​​and their indices to form the final lightweight model parameter set.

[0128] By combining sparsification and quantization compression, the local model parameter set updated in the previous step is transformed into a lightweight parameter set with lower computational overhead, higher storage efficiency, and preservation of key information, thus achieving the expected technical effect of balancing real-time performance and resource consumption in the online learning process.

[0129] For example, in a summer peak forecasting task for a regional power grid, a sparsity threshold is set. = Regularization coefficient = The updated weight matrix (1000×512) was compared element-by-element, and parameters with absolute values ​​less than a threshold were removed, achieving an initial sparsity of 72.3%. The gradient mean and variance were calculated, and parameters with importance scores below 0.05 were reset to zero, further increasing the sparsity to 78.5%. 8-bit uniform quantization was used, with the original floating-point parameter range [-0.12, 0.15] and a quantization step size of... Approximately After quantization, the storage size decreased from 2.048MB to 0.512MB. Using CSR format for storage, the compressed storage size is 0.11MB, accounting for 5.37% of the original size. The single prediction inference latency decreased from 12.8ms to 7.4ms, verifying that the method can significantly improve the efficiency of online updates and inference while maintaining prediction accuracy (accuracy decrease of less than 0.5%).

[0130] S5.4: Based on the compressed lightweight model parameter set and combined with the real-time influencing factors output by the dynamic weight adjustment module, the core parameters of the load forecasting model are replaced and fused to generate an online update model with the ability to adapt to external factors.

[0131] S5.5: Perform convergence verification and error assessment on the online update model. Based on the residual distribution characteristics between the current prediction output and the actual load data, determine whether the model meets the preset stability and accuracy requirements, so as to generate model state assessment results and trigger the subsequent parameter recalibration mechanism.

[0132] Step S6: Integrate the dynamic weight adjustment results and the updated parameters output by the online learning module into the load forecasting model to generate a load forecasting output with adaptive capabilities to external factors. Specifically, this includes: S6.1: Perform dimension alignment processing on the weight coefficients output by the dynamic weight allocation model and the model update parameters output by the online learning module to ensure that the input dimensions of the two are consistent in the load prediction model, thereby providing structural consistency guarantee for subsequent fusion calculation.

[0133] S6.2: Based on the weighted fusion strategy, the aligned dynamic weight coefficients are linearly combined with the model update parameters to generate a fusion parameter vector. The fusion coefficients are obtained by jointly optimizing the historical prediction error and the model stability index to balance prediction accuracy and model robustness.

[0134] The dynamic weight coefficient set after dimension alignment and the lightweight model update parameter set output by the online learning module are input into the fusion computing unit as the input vector for weighted fusion processing.

[0135] A linear weighted fusion algorithm is used (parameter: set of fusion coefficients). Number of features This enables the dynamic weight coefficient vector. With updating parameter vector The element-wise fusion is performed using the following formula: in For the fusion of the first Each parameter component For the corresponding fusion coefficient, and These are the dynamic weight coefficients and the updated model parameters, respectively. Each component.

[0136] Furthermore, through the fusion coefficient optimization module, a multi-objective joint optimization method is adopted (objective function: minimize historical prediction error). With predicted stability volatility The weighted sum (of the weighted sums) is calculated using the following formula: in The optimal balance coefficients are obtained iteratively on the training set using gradient descent. Sets to minimize .

[0137] Furthermore, constraints are introduced. This ensures that the fusion coefficient is a weighted proportional coefficient in a physical sense and avoids abnormal model parameters caused by numerical out-of-bounds errors.

[0138] By employing a weighted fusion strategy, optimized fusion coefficients are applied to the element-wise combination of dynamic weight coefficients and updated model parameters to generate a fusion parameter vector with consistent structure and balanced performance, thereby achieving simultaneous optimization of prediction accuracy and model robustness.

[0139] For example, in a coastal load forecasting task, the length of the dynamic weighting coefficient vector... = The element value is [ , , , , ]; Update the parameter vector to [ , , , , The initial fusion coefficient was calculated from historical data. , , , , Substitute into the fusion formula to calculate the first component: The fusion parameter vector is calculated sequentially as follows: , , , , After applying this vector to the load forecasting model, the mean absolute error of short-term forecasts was reduced in high load fluctuation scenarios during typhoons. Predicted volatility decreased This verifies the technical effectiveness of fusion coefficient optimization in improving the accuracy and stability of balance prediction.

[0140] S6.3: Input the fused parameter vector into the core calculation unit of the load forecasting model, replace the original model parameters, realize online parameter updates of the forecasting model, and enable the model to have the ability to respond instantly to changes in external factors.

[0141] S6.4: In the updated load forecasting model, forward forecasting calculations are performed based on multi-source external factor data within the current time window to generate short-term load forecasts with dynamic adaptability, and the results are output to the deviation analysis module for error assessment.

[0142] S6.5: Post-processing correction is performed on the generated load forecast values. The moving average filtering algorithm is used to smooth the forecast sequence to suppress short-term fluctuations in the parameter update process and improve the stability and engineering usability of the forecast output.

[0143] Step S7: Perform deviation analysis between the predicted output and the actual load data. If the deviation exceeds a preset threshold, trigger the model parameter recalibration mechanism to adjust the sliding window length and the correlation calculation weight factor. Specifically, this includes: S7.1: Compare the predicted load data output by the load forecasting model with the actual load data collected at the corresponding timestamp point by point to calculate the prediction deviation value at each time point.

[0144] S7.2: Based on the sliding window method, statistical analysis is performed on the prediction deviation values ​​of multiple consecutive time points to calculate the mean absolute error (MAE) and mean square error (MSE) within the current window, so as to quantify the overall prediction stability of the model.

[0145] S7.3: Compare the calculated mean absolute error with the preset error threshold. If the current error index continuously exceeds the threshold setting, it is determined that the model prediction performance has deteriorated, and the parameter recalibration mechanism is triggered.

[0146] The output of S7.2 is input into the mean absolute error calculation module to calculate the mean absolute error value within the current sliding time window. Error threshold set by the system Quantitative comparisons are performed to determine the real-time state of the model's predictive performance.

[0147] The error ratio calculation method is adopted (parameter: , To determine if an error exceeds the limit, the calculation formula is as follows: in Let this be the error ratio, if This indicates that the current prediction error exceeds the threshold limit.

[0148] Furthermore, through a continuous over-limit detection algorithm (parameter: number of consecutive over-limits), Monitoring cycle length This enables continuous monitoring of current error fluctuations, recording the out-of-limit state at each detection moment as a Boolean sequence. and in length of Count the number of times the limit is exceeded within the sliding window. .

[0149] Furthermore, threshold-triggered logic (parameter: , To determine model degradation, if and only if At that time, a performance degradation flag is generated. Otherwise set .

[0150] Furthermore, the performance degradation flag is input to the parameter recalibration trigger module. Then, a calibration trigger signal is sent to the subsequent sliding window adjustment and weight optimization steps to achieve adaptive adjustment of the load forecasting model.

[0151] By combining the out-of-limit characteristics of mean absolute error with time continuity through the above algorithm link, a precise trigger signal for performance degradation is generated, enabling the model to quickly self-correct when prediction accuracy declines.

[0152] For example, in a summer high-temperature load forecasting operation scenario, the following settings are configured: MW, the tolerance threshold for consecutive overruns Monitoring cycle length Each time interval (every 10 minutes) is calculated continuously within a monitoring cycle. The sequence is [ , , , , ]MW, the corresponding error ratio sequence is [ , , , , There were 3 instances of exceeding the limit during this period. ),satisfy The conditions are met, thus triggering the performance degradation flag. Subsequently, the sliding window length is shortened by 20%, and the relevance weights of external factors are reallocated. Ultimately, in subsequent prediction periods, [the following will be implemented / implemented]. The power consumption was reduced to 13.5 MW, ensuring the stability of the prediction model under continuous high temperature conditions.

[0153] S7.4: Based on the current error trend and historical error patterns, a fuzzy logic control algorithm is used to dynamically adjust the sliding time window length to enhance the model's sensitivity to short-term mutations and long-term trends.

[0154] S7.5: Based on the adjusted sliding window length, recalculate the dynamic correlation coefficients between each external factor and load change, and optimize the weight factor allocation strategy in the correlation calculation based on the weighted least squares method.

[0155] Step S8: Based on historical prediction errors and model recalibration response data, generate model performance logs and update the knowledge base for modeling the impact of external factors, which is used for initial configuration optimization in subsequent prediction cycles. Specifically, this includes: S8.1: Classify and statistically analyze historical prediction error data, extract error distribution characteristics, and identify the performance degradation mode of the model under different combinations of external factors.

[0156] S8.2: Based on the model recalibration response data, analyze the response pattern of sliding window length adjustment and correlation calculation weight factor change, so as to quantify the impact of dynamic changes in external factors on the adaptability of model parameters.

[0157] S8.3: Perform cross-correlation analysis between error distribution characteristics and parameter adjustment response patterns to generate a structured model performance log, recording prediction error types, calibration actions, and their corresponding combinations of external factors.

[0158] Based on the error distribution characteristics extracted in S8.1 and the parameter adjustment response law output in S8.2, the correlation coefficient analysis method is adopted (parameter: error type code E). i Adjusting action code A j External factor combination vector F k This allows for the preliminary calculation of the correlation between different error types and parameter adjustment actions.

[0159] Furthermore, through mutual information calculation methods (parameter: E) i A j This allows for the quantification of the correlation between error types and adjustment actions under a joint probability distribution, and yields an initial correlation matrix reflecting the combination of sensitive factors. .

[0160] Furthermore, based on the multiple regression analysis method (parameter: F... k E i A j This allows for the estimation of the influence coefficients of external factor combinations on specific error types and their triggered parameter adjustment actions, and the generation of a normalized influence coefficient matrix. .

[0161] Furthermore, through a matrix fusion algorithm (parameters: , The weighted synthesis of correlation and influence coefficients (α) is achieved to obtain the structured model performance log matrix. The fusion formula is as follows: By using field mapping rules (parameters: error type, adjustment action, combination of external factors), The matrix is ​​transformed record by record into a structured model performance log containing error type identifiers, corresponding parameter calibration actions, and information on the combination of related external factors.

[0162] Through the aforementioned algorithmic chain, independent error distribution characteristics and parameter adjustment response patterns are cross-correlated and analyzed to generate a data foundation that can be used for knowledge graph construction, thereby realizing the transformation of historical performance experience into queryable and reasonable structured data.

[0163] For example, in a scenario of fluctuating electricity prices during low temperatures in winter, the error type code E i Including two categories: "systematic overestimation" and "random underestimation," the parameter adjustment action code A j The external factor combination vector F includes two categories: "shortening the sliding window" and "increasing the weight factor". k This includes a combination of three factors: a sudden drop in temperature, a sudden increase in electricity prices, and changes in residential electricity consumption patterns. E was calculated using correlation coefficient analysis. i With A j Preliminary correlation matrix between The matrix elements are in the range [0,1], such as =0.82. The relationship between F_k and observed E is obtained through multiple regression. i and A j The influence coefficient matrix of the combination ,like =0.76. Let the fusion weight coefficient be... =0.6, then the fusion calculation formula is: Finally, the structured model performance log generated an entry for "systematic overestimation - increased weight factor - sudden drop in temperature + sudden increase in electricity price", with an intensity index of 0.796, which was used to support the optimization of prediction initialization parameter settings in the subsequent knowledge graph construction stage.

[0164] S8.4: Based on the performance log of the structured model, construct a knowledge graph representation of the influence pattern of external factors, and use graph database technology to establish semantic associations between factors, errors and calibration actions.

[0165] S8.5: Write the updated knowledge graph data into the knowledge base for modeling the influence of external factors, and provide initial configuration suggestions for the next prediction cycle, including the initial sliding window length, correlation weight factors and prior distribution of dynamic weight allocation.

[0166] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the element or object preceding "comprising" or "including" covers the element or object listed following "comprising" or "including" and its equivalents, and does not exclude other elements or objects. The multiple involved in the embodiments of this application refers to two or more. A and / or B indicate the presence of three cases: A; B; and A and B.

[0167] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic monitoring and evaluation of load demand in a smart grid, specifically comprising: S1: Deploy a multi-source external factor sensing network to collect multi-dimensional external factor data and record the data collection timestamp to achieve time-series alignment; S2: Preprocess the collected multi-source external factor data; S3: Based on the sliding time window, perform dynamic correlation analysis on the normalized external factor data and historical load data, and calculate the real-time correlation coefficient matrix between each external factor and load change; S4: Construct a dynamic weight allocation model based on the correlation coefficient matrix, and dynamically adjust the input weights of different external factors in the load forecasting model according to the influence intensity of different external factors in the current time window; S5: Input the weighted multi-source external factor data into the lightweight online learning module, and use the incremental learning algorithm to update the load forecasting model parameters in real time without retraining the global model; S6: Integrate the dynamic weight adjustment results with the updated parameters output by the online learning module into the load forecasting model to generate a load forecasting output with adaptive capabilities to external factors.

2. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, Step S6 is followed by: S7: Perform deviation analysis between the predicted output and the actual load data. If the deviation exceeds the preset threshold, trigger the model parameter recalibration mechanism to adjust the sliding window length and correlation calculation weight factor. S8: Based on historical prediction errors and model recalibration response data, generate model performance logs and update the knowledge base for modeling the impact of external factors, which is used for initial configuration optimization in subsequent prediction cycles.

3. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, The multidimensional external data includes meteorological parameters, holiday types, real-time electricity prices, and user behavior characteristics.

4. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, Preprocessing of the collected multi-source external factor data includes normalization and outlier filtering to eliminate the impact of different dimensions and sudden noise on subsequent modeling.

5. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 3, characterized in that, Communication protocol adaptation configuration is performed on meteorological monitoring stations, electricity price release systems, user-side smart meters, and behavior collection terminals deployed within the target power grid area to support unified access to multi-source heterogeneous data.

6. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, In step S1, timestamp annotation is performed on the collected external factor data, and the time base of each data source is uniformly calibrated using the NTP synchronization mechanism to achieve time sequence consistency alignment of cross-source data.

7. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, Step S3 specifically includes: The normalized external factor data and historical load data are time-aligned, and the data sequence is divided into units of sliding time windows to generate a time series data subset for correlation analysis. For each sliding window, the pairwise correlation between the external factor data subset and the corresponding load data subset is calculated to obtain the initial correlation coefficient matrix; A dynamic weighting factor is introduced to weight the initial correlation coefficient matrix. The dynamic weighting factor is set to decay exponentially according to the sliding direction of the time window to generate a dynamically adjusted correlation coefficient matrix. Based on the dynamically adjusted correlation coefficient matrix, the high-dimensional external factors are reduced in dimensionality to extract the main influencing factors and their corresponding contribution indicators. The principal component analysis results are cross-validated with the load change trend to generate a dynamic influence map between external factors and load changes.

8. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, Step S4 specifically includes: The real-time correlation coefficient matrix of each external factor and load change calculated within the sliding time window is normalized to obtain the normalized correlation coefficient vector. Based on the normalized correlation coefficient vector, an exponential decay function is used to amplify the weights of highly correlated external factors to enhance their dynamic influence weight in the load forecasting model, thereby obtaining a dynamic weight amplification coefficient matrix. A time decay factor is introduced to adjust the time sensitivity of the dynamic weight amplification coefficient matrix, generating a time-weighted dynamic weight matrix. Based on the time-weighted dynamic weight matrix, a multi-factor dynamic weight allocation model is constructed. The dynamic weights of each external factor are mapped to the input feature space of the load forecasting model through a linear weighted fusion method, generating a weighted input feature vector. The weighted input feature vector is output to a lightweight online learning module as the input parameters for the load forecasting model.

9. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 1, characterized in that, The model performance log in S8 includes error type, parameter calibration action, and external factor combination triplet.

10. The method for dynamic monitoring and evaluation of smart grid load demand according to claim 9, characterized in that, Influence weights are obtained through association analysis and multiple regression, and logs are entered into an external factor knowledge base to generate a knowledge graph.

Images