A power distribution network dispatching operation situation prediction method, system, device and medium

Abstract

The application discloses a power distribution network dispatching operation situation prediction method, system, equipment and medium, relates to the power system operation technical field, and comprises the following steps: an attention mechanism is used to allocate initial weights to each source information of the power distribution network, and weighted fusion features are obtained by weighting and fusing the multi-source information; time sequence correlation information of each time step in the weighted fusion features is extracted to obtain a hidden state sequence; abnormal detection is performed on the hidden state sequence, when it is detected that the sequence value deviates from a preset range, the local fluctuation amplitude of the sequence in the window is extracted, and an attention weight matrix is updated to generate a dynamic contribution degree vector; the weights of each source information in the weighted fusion features are adjusted according to the dynamic contribution degree vector, the network parameters are updated according to the adjusted weights to obtain an optimized prediction model; the real-time collected multi-source information is input into the optimized prediction model, a prediction result is obtained, and a situation prediction result is output according to the prediction result. The application solves the problems of poor prediction accuracy and dispatching reliability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of power system operation technology, and in particular to a method, system, equipment and medium for predicting the status of power distribution network dispatching operations. Background Technology

[0002] As a crucial link in the power system directly facing end users, the accuracy and timeliness of the power distribution network's dispatching operations directly affect power supply reliability, grid operation safety, and the normal operation of socio-economic activities. With the continuous growth of societal demand for electricity and the in-depth advancement of new power system construction, a large number of distributed power sources, energy storage devices, and diverse loads are widely integrated, leading to increasingly complex distribution network structures and flexible operation modes, significantly increasing the uncertainties faced by dispatching operations. Against this backdrop, distribution network dispatching situation prediction—that is, comprehensively predicting the distribution network's operating status, operational risks, and environmental impacts within a specific time window—has become one of the key technologies for ensuring the safe and stable operation of the power grid. Accurate situation prediction not only relates to the efficient operation of the power grid but also plays a decisive role in rapid response to emergencies and the rational allocation of resources.

[0003] Currently, researchers have conducted extensive work on predicting the operational status of power distribution networks. Existing methods typically collect equipment operating status and surrounding environmental parameters from various sensors deployed on power transmission lines (such as wind speed sensors, temperature sensors, and humidity sensors), while simultaneously combining this data with regional weather data transmitted from meteorological stations (such as wind direction and atmospheric pressure), forming a multi-source information input. Based on this, mathematical statistics methods or simple weighted fusion strategies are used to integrate the multi-source data, and then combined with historical records for situation analysis. Some more advanced solutions introduce deep learning models such as Long Short-Term Memory (LSTM) networks, utilizing their ability to model time-series data to extract temporal correlation features to predict future states. These methods can provide a certain degree of reference for dispatching operations under relatively stable operating environments.

[0004] However, current research methods often suffer from the following major problems when dealing with complex environments. First, the weight allocation in multi-source information fusion lacks dynamic adaptability. Existing methods, when fusing heterogeneous data from multiple sources such as sensors and weather stations, typically employ fixed weights or static weighting strategies based on experience, failing to reasonably balance the reliability differences between data from different sources. The characteristics and importance of various types of data differ greatly; for example, the impact of equipment operation records and weather change information on prediction varies. If the contribution of each information source cannot be dynamically adjusted according to environmental changes and fluctuations in data quality, it can lead to the omission or interference of key information, affecting prediction accuracy. For instance, in a stormy weather event, if the importance of meteorological data is not fully considered and static weights are used for fusion, it may lead to an underestimation of the risk of line faults, ultimately causing local power outages and dispatch delays. Second, the predictive models lack the ability to respond to and adaptively adjust to dynamic environmental changes. Most existing prediction methods adopt a fixed pattern of offline training and online application, with model parameters relatively fixed after deployment. When the operating environment of the distribution network undergoes significant changes (such as sudden extreme weather or equipment failure leading to topology changes), existing solutions struggle to adjust forecasting strategies in a timely manner. They fail to effectively capture the state transition patterns caused by sudden environmental changes, resulting in forecasts that often deviate from actual operating conditions and impacting the accuracy of dispatching decisions. This lag is particularly pronounced in the face of emergencies, making it difficult for the system to quickly adapt to new trends and weakening the accuracy and reliability of distribution network dispatching operations in complex and volatile environments. In summary, existing technologies employ static weights and fixed model parameters in multi-source information fusion, failing to dynamically adapt to environmental changes, leading to poor forecasting accuracy and dispatching reliability. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for predicting the status of power distribution network dispatching operations. It addresses the problem that existing technologies use static weights and fixed model parameters in multi-source information fusion, which cannot dynamically adapt to environmental changes, resulting in poor prediction accuracy and dispatching reliability.

[0006] In view of this, the first aspect of the present invention provides a method for predicting the status of power distribution network dispatching operations, the method comprising: Multi-source information transmitted by sensors and weather stations in the power distribution network is collected. An attention mechanism is used to assign initial weights to each source of information and construct an attention weight matrix. The attention weight matrix is ​​then used to perform weighted fusion of the multi-source information to obtain weighted fusion features. The weighted fusion features are input into a long short-term memory network, and the temporal correlation information of each time step is extracted through the gating mechanism of the long short-term memory network to obtain the hidden state sequence. Anomaly detection is performed on the hidden state sequence. When the sequence value deviates from the preset range, sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The attention weight matrix is ​​updated based on the local fluctuation amplitude to generate a dynamic contribution vector. The weights of each source information in the weighted fusion feature are adjusted according to the dynamic contribution vector. The weighted fusion feature with adjusted weights is then input into the long short-term memory network for secondary training to update the network parameters and obtain the optimized prediction model. The multi-source information collected in real time is input into the optimized prediction model to obtain the prediction results. The environmental change trend is determined based on the prediction results. The state probability distribution is adjusted according to the environmental change trend. Based on the adjusted state probability distribution and historical records, the key correlation path is strengthened through an attention mechanism, and the situation prediction result is output.

[0007] Optionally, the process of collecting multi-source information transmitted from sensors and weather stations in the power distribution network employs an attention mechanism to assign initial weights to each source of information and construct an attention weight matrix. The attention weight matrix is ​​then used to perform weighted fusion of the multi-source information to obtain weighted fusion features, including: Environmental data is collected by sensors deployed in the power distribution network, and real-time meteorological indicators are obtained from meteorological stations to form a multi-source information sequence; The multi-source information sequence is preprocessed to convert the information from each source into a feature vector in numerical form; A query vector is generated based on the current prediction target. The feature vectors of each source information are input into the attention mechanism to generate corresponding key vectors. The dot product similarity between the query vector and the key vectors corresponding to each source information is calculated. After scaling, the similarity is normalized by the softmax function to obtain the attention score of each source information as the initial weight. An attention weight matrix is ​​constructed based on the initial weights of each source information. The feature vectors of each source information are then weighted and summed using the attention weight matrix to obtain the weighted fusion feature.

[0008] Optionally, the step of inputting the weighted fusion features into a long short-term memory network and extracting the temporal correlation information of each time step through the gating mechanism of the long short-term memory network to obtain the hidden state sequence includes: The weighted fusion features are arranged according to time steps to construct a time-series input sequence; The time-series input sequence is sequentially input into the Long Short-Term Memory (LSTM) network. At each time step, the input gate controls the amount of new information written, the forget gate controls the amount of cell state retained from the previous time step, and the output gate controls the amount of hidden state output at the current time step, thus updating the cell state and hidden state. The hidden state output at each time step is recorded to form a hidden state sequence.

[0009] Optionally, the anomaly detection of the hidden state sequence, when a sequence value deviates from a preset range, involves initiating a sliding window analysis to extract the local fluctuation amplitude of the sequence within the window, including: Obtain the hidden state value at each time step in the hidden state sequence, and calculate the difference in hidden state values ​​between adjacent time steps; The difference is compared with a preset threshold. When the difference exceeds the preset threshold, it is determined that the current sequence value deviates from the preset range and a sliding window analysis is started. The window length of the sliding window is determined, and the window is slid on the hidden state sequence with a fixed step size to extract the hidden state subsequence within the current window. Calculate the mean and variance of each hidden state value in the hidden state subsequence, and determine the local fluctuation amplitude within the current window based on the mean and variance.

[0010] Optionally, updating the attention weight matrix based on the local fluctuation amplitude to generate a dynamic contribution vector includes: The weight adjustment factor corresponding to each source of information is calculated based on the local fluctuation amplitude within the current window, and the weight adjustment factor is positively correlated with the local fluctuation amplitude; The weight adjustment factor corresponding to each source information is multiplied by the initial weight at the corresponding position in the attention weight matrix to obtain the updated attention weight matrix; The updated attention weight matrix is ​​weighted and calculated with the feature vectors of each source information in the weighted fusion feature to generate a dynamic contribution vector.

[0011] Optionally, adjusting the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and inputting the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters and obtain an optimized prediction model, includes: The current weight of each source information in the weighted fusion feature is added to the element value at the corresponding position in the dynamic contribution vector to obtain the adjusted weight, wherein each element in the dynamic contribution vector corresponds to the weight adjustment value of each source information in the weighted fusion feature. The multi-source information is re-weighted and fused using the adjusted weights to obtain the weighted fusion features with adjusted weights. The weighted fusion features after weight adjustment are used as training samples and input into the long short-term memory network. The network parameters of the long short-term memory network are updated through the backpropagation algorithm. Calculate the loss function value of the Long Short-Term Memory network. If the loss function value satisfies the preset convergence condition, stop training and output the current network parameters as the optimized prediction model. If the loss function value does not satisfy the preset convergence condition, continue to perform backpropagation to update the network parameters and repeat training until the preset convergence condition is satisfied.

[0012] Optionally, the step of inputting real-time collected multi-source information into the optimized prediction model to obtain prediction results, determining the environmental change trend based on the prediction results, adjusting the state probability distribution based on the environmental change trend, strengthening key correlation paths through an attention mechanism based on the adjusted state probability distribution and historical records, and outputting the situation prediction result includes: The multi-source information collected in real time is input into the optimized prediction model, and the prediction result is output after forward calculation of the model. The prediction results are subjected to time series analysis to calculate the rate of change of each environmental parameter. The rate of change is then compared with a preset trend threshold to determine the current environmental change trend. A state transition matrix is ​​constructed based on the environmental change trend, and the adjusted state probability distribution is calculated using the Markov chain method based on the state transition matrix. The adjusted state probability distribution is compared with the state probability vectors at each time point in the historical record to calculate the matching degree, and attention weights are assigned to each historical record based on the matching degree. Based on the attention weights, relevant state sequences are extracted from historical records to construct focus-related paths; path dependency modeling is performed on the focus-related paths to extract the state transition relationships of each associated node in the path, and the state transition relationships are dynamically corrected by combining real-time multi-source information to obtain situation prediction results.

[0013] A second aspect of the present invention provides a distribution network dispatching operation status prediction system, the system comprising: The construction unit is used to collect multi-source information transmitted by sensors and weather stations in the power distribution network, and uses an attention mechanism to assign initial weights to each source of information and construct an attention weight matrix. The attention weight matrix is ​​then used to perform weighted fusion of the multi-source information to obtain weighted fusion features. The extraction unit is used to input the weighted fusion features into the long short-term memory network, and extract the temporal correlation information of each time step through the gating mechanism of the long short-term memory network to obtain the hidden state sequence; The generation unit is used to perform anomaly detection on the hidden state sequence. When the sequence value is detected to deviate from the preset range, a sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The attention weight matrix is ​​updated according to the local fluctuation amplitude to generate a dynamic contribution vector. The training unit is used to adjust the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and input the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters and obtain the optimized prediction model. The output unit is used to input multi-source information collected in real time into the optimized prediction model, obtain prediction results, determine the environmental change trend based on the prediction results, adjust the state probability distribution based on the environmental change trend, strengthen key correlation paths through an attention mechanism based on the adjusted state probability distribution and historical records, and output the situation prediction result.

[0014] A third aspect of the present invention provides a distribution network dispatching operation status prediction device, the device comprising a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the steps of the distribution network dispatching operation status prediction method as described in the first aspect above, according to the instructions in the program code.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium for storing program code for executing the distribution network dispatching operation status prediction method described in the first aspect above.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a method for predicting the status of power distribution network dispatching operations. Addressing the challenges of multi-source heterogeneous data fusion, delayed response to abnormal fluctuations, and insufficient adaptability of prediction models to dynamic environments in power distribution network dispatching operations, the method first collects multi-source information transmitted from sensors and weather stations in the power distribution network and assigns initial weights to each source using an attention mechanism. Weighted fusion of the multi-source information yields a weighted fusion feature, allowing the initial weights to be dynamically quantified based on the correlation between each source and the prediction target, avoiding the omission or interference of key information caused by fixed weight allocation. Next, the weighted fusion feature is input into a Long Short-Term Memory (LSTM) network, and its gating mechanism extracts the temporal correlation information at each time step to obtain a hidden state sequence. This allows the network to selectively retain or discard information, achieving accurate representation of temporal dependencies. Finally, when a hidden state sequence value deviates from a preset range, a sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The system dynamically updates the attention weight matrix based on local fluctuations to generate a dynamic contribution vector, enabling the attention weight matrix to adaptively update according to the degree of environmental fluctuations, thus solving the problem of static weights lacking dynamic adaptability. Then, the weights of each source of information are adjusted according to the dynamic contribution vector, and the weighted fusion features with adjusted weights are input again into the Long Short-Term Memory (LSTM) network for secondary training to update the network parameters, resulting in an optimized prediction model. This optimizes the model parameters to adapt to the current abnormal fluctuation environment, solving the problem of delayed response to environmental changes caused by fixed model parameters. Finally, real-time multi-source information is input into the optimized prediction model to obtain prediction results. Based on the prediction results, the environmental change trend is determined to adjust the state probability distribution. Based on the adjusted state probability distribution and historical records, key correlation paths are strengthened through an attention mechanism to output situational prediction results, ensuring that the prediction output takes into account both the reference value of historical experience and the dynamic changes of the real-time environment. The entire solution, through the synergistic mechanism of dynamic weight adjustment and model parameter optimization, solves the technical problems of insufficient prediction accuracy and scheduling reliability caused by static weights in multi-source information fusion and fixed model parameters, significantly improving the situational awareness accuracy and anomaly response speed of distribution network scheduling in complex and ever-changing environments. In summary, this invention solves the problem that existing technologies use static weights and fixed model parameters in multi-source information fusion, which cannot dynamically adapt to environmental changes, resulting in poor prediction accuracy and scheduling reliability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1A flowchart illustrating a method for predicting the status of power distribution network dispatching operations, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a power distribution network dispatching operation status prediction system provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] Please see Figure 1 The present invention provides a method for predicting the status of power distribution network dispatching operations, comprising: Step 101: Collect multi-source information transmitted by sensors and weather stations in the power distribution network, assign initial weights to each source of information using an attention mechanism and construct an attention weight matrix, and use the attention weight matrix to perform weighted fusion of the multi-source information to obtain weighted fusion features; Step 102: Input the weighted fusion features into the Long Short-Term Memory (LSTM) network, and extract the temporal correlation information of each time step through the gating mechanism of the LTM network to obtain the hidden state sequence; Step 103: Perform anomaly detection on the hidden state sequence. When the sequence value is detected to deviate from the preset range, start sliding window analysis to extract the local fluctuation amplitude of the sequence within the window; update the attention weight matrix according to the local fluctuation amplitude to generate a dynamic contribution vector. Step 104: Adjust the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and input the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters and obtain the optimized prediction model. Step 105: Input the real-time collected multi-source information into the optimized prediction model to obtain the prediction results. Determine the environmental change trend based on the prediction results. Adjust the state probability distribution based on the environmental change trend. Based on the adjusted state probability distribution and historical records, strengthen the key correlation path through an attention mechanism and output the situation prediction results.

[0021] In one embodiment, step 101 includes: Step 1011: Collect environmental data through sensors deployed in the power distribution network and obtain real-time meteorological indicators from the meteorological station to form a multi-source information sequence; First, environmental data is collected by sensors deployed in the power distribution network. Specifically, these sensors include wind speed sensors, temperature sensors, and humidity sensors, which are installed at key nodes of the distribution network (such as near transmission towers or conductors) to monitor micro-meteorological parameters around the conductors in real time. These parameters include wind speed, temperature readings, and humidity percentages, and the data is transmitted to the central processing system via a wireless network at fixed intervals (e.g., once per minute). Then, real-time meteorological indicators are obtained from weather stations, i.e., regional weather data, such as wind direction and atmospheric pressure, are received via API interfaces. Thus, the precise local data from the sensors and the macro-trend data from the weather stations are combined to form a multi-source information sequence. This sequence encompasses multi-dimensional state information of the equipment's operating environment, providing a comprehensive and synchronous data foundation for subsequent attention mechanisms to perform adaptive weight allocation and feature fusion.

[0022] Step 1012: Preprocess the multi-source information sequence to convert the information from each source into a feature vector in numerical form; First, the multi-source information sequence constructed in step 1011 is preprocessed. Since the raw data collected by sensors and weather stations may contain noise or abnormal fluctuations during transmission (e.g., large jumps in wind speed sensor readings due to instantaneous gusts), the raw data sequence needs noise removal. Specifically, an averaging filter method can be used to smooth out outliers, reducing the impact of transient interference on data quality and ensuring the stability and reliability of the input data. Then, the preprocessed information from each source is uniformly converted into numerical feature vectors. For example, physical quantities such as wind speed (e.g., 15.0 m / s), temperature reading (e.g., 25.5℃), and humidity percentage (e.g., 80%) are encoded into multi-dimensional numerical vectors. Thus, the raw collected data is transformed into standardized numerical inputs that can be directly processed by subsequent attention mechanisms and long short-term memory networks, effectively reducing the interference of dimensional differences and noise on model training, and improving the accuracy of feature fusion and the stability of model training.

[0023] Step 1013: Generate a query vector based on the current prediction target, input the feature vectors of each source information into the attention mechanism to generate corresponding key vectors, calculate the dot product similarity between the query vector and the key vectors corresponding to each source information, and normalize it through the softmax function after scaling to obtain the attention score of each source information as the initial weight. First, a query vector is generated based on the current prediction target. The numerical feature vectors corresponding to each source information obtained in step 1012 are input into the attention mechanism to generate key vectors for each source information (e.g., wind speed feature vector generates a wind speed key vector, and temperature feature vector generates a temperature key vector). Then, the dot product similarity between the query vector and the key vectors corresponding to each source information is calculated. This dot product value reflects the correlation strength between each source information and the current prediction target. After aggregating the dot product similarity results of all source information, the results are scaled to prevent gradients from entering the saturation region, and then normalized using the softmax function to convert them into probability distribution weight values, obtaining the attention score for each source information. This attention score is then used as the initial weight for each source information. This achieves a dynamic quantitative assessment of the importance of each source information, allowing information with higher relevance to the distribution network situation prediction target (such as wind speed data during strong winds) to receive higher weights in subsequent fusion processes. This highlights the contribution of key information to the prediction results and avoids the information imbalance problem caused by fixed weight allocation.

[0024] Step 1014: Construct an attention weight matrix based on the initial weights of each source information, and use the attention weight matrix to perform a weighted summation of the feature vectors of each source information to obtain a weighted fusion feature.

[0025] First, based on the initial weights of each source information obtained in step 1013, these weights are organized according to the feature dimensions corresponding to each source information to construct an attention weight matrix. Each element in this matrix corresponds to the weight coefficient of the feature vector of each source information, used to characterize the relative importance of each source information in the current prediction scenario. Then, the feature vectors of each source information are weighted and summed using the attention weight matrix, that is, the feature vector of each source information is multiplied by the weight coefficient corresponding to that source information, and then the weighted results of all source information are added element by element to merge into a unified weighted fusion feature vector (for example, in stormy weather, the wind speed feature vector is assigned a higher weight due to the attention mechanism, and its information dominates in the fusion feature after weighted summation, while features with gradual changes, such as temperature, have a lower proportion). Thus, adaptive fusion of multi-source heterogeneous information is realized, integrating scattered and multidimensional original data into a compact and information-rich fusion feature representation, providing a high-quality input foundation for subsequent temporal modeling of long short-term memory networks.

[0026] In one embodiment, step 102 includes: Step 1021: Arrange the weighted fusion features according to time steps to construct a time-series input sequence; First, the weighted fusion feature obtained in step 1014 is acquired. This feature is a comprehensive representation after adaptive fusion of multi-source information. However, since it is still a static feature vector of a single time segment, it cannot reflect the dynamic evolution of the distribution network's operating status. Then, according to the time sequence of sensor and weather station data acquisition, the weighted fusion features of multiple consecutive time steps are arranged by time step. That is, each time step corresponds to a weighted fusion feature vector at a certain acquisition time. These vectors are organized into a time-series input sequence according to their chronological order (for example, the feature vectors collected and fused every minute in the past T time steps are arranged sequentially to form a time series of length T). Thus, the discrete static fusion features are organized into a serialized data structure containing a time dimension, providing an input format that conforms to the processing logic of the subsequent Long Short-Term Memory network to extract the temporal correlation information between each time step, enabling the network to capture the changing patterns of the distribution network's operating status along the time axis.

[0027] Step 1022: Input the time-series input sequence into the Long Short-Term Memory network in sequence. At each time step, control the amount of new information written through the input gate, control the amount of cell state retained from the previous time step through the forget gate, and control the amount of hidden state output at the current time step through the output gate to update the cell state and hidden state; record the hidden state output at each time step to form a hidden state sequence.

[0028] First, the temporal input sequence constructed in step 1021 is sequentially input into a long short-term memory network with a defined structure. The core of this network lies in its internal gating mechanism. At each time step, the network performs the following update operations: the input gate controls the amount of new information written, i.e., the input gate determines how much new information in the input features of the current time step is allowed to be added to the cell state; the forget gate controls the amount of cell state retained from the previous time step, i.e., the forget gate determines how much historical information in the cell state of the previous time step needs to be retained or discarded; and the output gate controls the amount of hidden state output at the current time step, i.e., the output gate determines how much information in the updated cell state is output as the hidden state of the current time step. The three mechanisms work together to update the cell state and hidden state (for example, when the wind speed data changes drastically at a certain time step, the input gate increases the proportion of new information written, while the forget gate appropriately reduces the proportion of steady-state weather information retained from the previous time step, allowing the cell state to respond quickly to environmental changes). Then, the hidden states output at each time step are recorded, and these hidden states are aggregated in chronological order to form a hidden state sequence. Thus, the Long Short-Term Memory Network (LSTM) achieves long-term memory of effective information and dynamic forgetting of irrelevant information in the temporal input sequence through its gating mechanism. This allows the generated hidden state sequence to retain key temporal dependencies while filtering out interference from redundant information, providing an accurate state representation basis for subsequent anomaly detection and dynamic weight updates.

[0029] In one embodiment, step 103 includes: Step 1031: Obtain the hidden state value at each time step in the hidden state sequence, and calculate the difference in hidden state values ​​between adjacent time steps; First, the hidden state values ​​for each time step are obtained from the hidden state sequence obtained in step 1022. Since each time step in the hidden state sequence corresponds to the encoding vector output by the Long Short-Term Memory network at that time step, this vector numerically summarizes the temporal correlation information up to the current time step. Then, the difference in hidden state values ​​between adjacent time steps is calculated point by point on the time series. This difference reflects the degree of change in the operating state of the distribution network between adjacent time points (for example, when the hidden state value of an adjacent time step jumps from 0.2 to 0.8, the difference is 0.6, indicating that the system state has changed significantly during this period). Thus, the calculation of the difference provides a quantifiable basis for subsequent anomaly detection, enabling the objective identification of time points in the hidden state sequence that deviate from the normal operating mode.

[0030] Step 1032: Compare the difference with a preset threshold. When the difference exceeds the preset threshold, determine that the current sequence value deviates from the preset range and start sliding window analysis. Determine the window length of the sliding window, slide the window on the hidden state sequence with a fixed step size, and extract the hidden state subsequence within the current window. Specifically, firstly, the difference calculated in step 1031 is compared with a preset threshold, which is used to distinguish between normal operation fluctuations and abnormal state abrupt changes. When the difference exceeds the preset threshold (for example, in normal conditions, the difference between hidden state values ​​of adjacent time steps is usually within 0.1, while the current difference reaches 0.6, exceeding the preset threshold of 0.3), it is determined that the current sequence value deviates from the preset range, indicating that the operating state near this time point has experienced significant fluctuations, requiring in-depth analysis. Then, a sliding window analysis is initiated, and the window length is determined (for example, the window length is fixed at 10 time steps). The window is slid across the hidden state sequence with a fixed step size, that is, a subsequence of continuous window length including the anomaly point is extracted forward from the anomaly point, and the hidden state subsequence within the current window is extracted. Thus, the sliding window analysis transforms the anomaly determination result into a specific local data extraction operation. The extracted hidden state subsequence gathers the state information of the anomaly point and its surrounding time steps, providing a focused data range for the subsequent accurate calculation of local fluctuation amplitude.

[0031] Step 1033: Calculate the mean and variance of each hidden state value in the hidden state subsequence, and determine the local fluctuation amplitude within the current window based on the mean and variance; First, the hidden state subsequence within the current window captured in step 1032 is obtained. Statistical calculations are performed on the hidden state values ​​corresponding to each time step within this subsequence, calculating the mean and variance of all hidden state values ​​within the subsequence. The mean reflects the overall level of hidden states within the window, while the variance reflects the dispersion of hidden state values ​​at each time step relative to the mean. Then, the local fluctuation amplitude within the current window is determined based on the mean and variance (for example, a large variance indicates drastic state changes between time steps within the window, resulting in correspondingly large local fluctuations; a small variance indicates smooth state changes within the window, resulting in smaller local fluctuations). Thus, the mean and variance, as two core statistical measures of data distribution within the window, quantitatively describe the drastic fluctuation of hidden states within the window from two dimensions: central tendency and dispersion. This provides a quantitative input basis for the subsequent calculation of the weight adjustment factor, enabling windows with greater fluctuations to trigger more significant weight adjustments.

[0032] Step 1034: Calculate the weight adjustment factor corresponding to each source information based on the local fluctuation amplitude within the current window, wherein the weight adjustment factor is positively correlated with the local fluctuation amplitude; First, the local fluctuation amplitude within the current window determined in step 1033 is obtained. This local fluctuation amplitude is a quantitative representation of the intensity of fluctuation in the hidden state sequence within the window, reflecting the instability of the distribution network operation state within the window period. Then, the weight adjustment factor corresponding to each source information is calculated based on the local fluctuation amplitude. Since the weight adjustment factor is positively correlated with the local fluctuation amplitude, that is, the larger the local fluctuation amplitude, the larger the weight adjustment factor corresponding to each source information, indicating that the current environment changes more drastically, and the more the source information needs to be rebalanced during the fusion process (for example, when the variance within a certain window is large, resulting in a local fluctuation amplitude of 0.8, the calculated weight adjustment factor is 1.6, while the weight adjustment factor corresponding to a local fluctuation amplitude of 0.3 is only 1.1). Thus, the weight adjustment factor transforms the degree of abnormal fluctuation into a quantifiable weight correction coefficient, providing a basis for dynamic adjustment for subsequent updates to the attention weight matrix, so that the weight adjustment amplitude of each source information is larger in scenarios with greater fluctuations.

[0033] Step 1035: Multiply the weight adjustment factor corresponding to each source information with the initial weight at the corresponding position in the attention weight matrix to obtain the updated attention weight matrix; First, the weight adjustment factors corresponding to each source information calculated in step 1034 and the initial weights corresponding to each source information in the attention weight matrix constructed in step 1014 are obtained. Each element in the attention weight matrix corresponds to the initial weight value of the feature vector of each source information. Then, the weight adjustment factors corresponding to each source information are multiplied by the initial weights at the corresponding positions in the attention weight matrix (for example, the initial weight of wind speed source information is 0.5, and its corresponding weight adjustment factor is 1.6, so it is updated to 0.8 after multiplication; the initial weight of temperature source information is 0.3, and its corresponding weight adjustment factor is 1.1, so it is updated to 0.33 after multiplication). After the weights at the corresponding positions of all source information are multiplied in the above way, the updated attention weight matrix is ​​obtained. Thus, the original weights of each source information are adaptively amplified or reduced according to the degree of environmental fluctuation, realizing the dynamic updating of the attention weight matrix. This allows the updated attention weight matrix to reflect the true importance distribution of each source information in the current environment, providing a weight basis for generating dynamic contribution vectors.

[0034] Step 1036: Perform a weighted operation on the updated attention weight matrix and the feature vectors of each source information in the weighted fusion feature to generate a dynamic contribution vector.

[0035] First, obtain the updated attention weight matrix obtained in step 1035 and the feature vectors of each source information obtained in step 1012. Each element in the updated attention weight matrix corresponds to the updated weight of each source information, and the feature vector is a numerical vector obtained by preprocessing each source information. Then, perform a weighted operation on the updated attention weight matrix and the feature vectors of each source information in the weighted fusion feature. That is, for each source information, multiply the value of each dimension of the feature vector of that source information by the updated weight corresponding to that source information, and then combine the weighted results of all source information (for example, the updated weight of wind speed source information is 0.8, and its corresponding feature vector contains a wind speed value of 15.0). If the temperature source information has a weight of 0.4 after the weighted operation (the corresponding component in its contribution vector is 12.0), and its feature vector contains a temperature value of 25.5℃, then the corresponding component in its contribution vector after the weighted operation is 10.2, the final result is a dynamic contribution vector. Each element in this vector quantifies the dynamic contribution of each source information to the fusion feature under the current distribution network operating environment. Thus, the dynamic contribution vector effectively combines the updated weight information with the original features of each source information, quantifying the real-time contribution of each source information to the fusion feature under abnormal fluctuation conditions, providing direct data support for subsequent adjustments to the weights of each source information in the weighted fusion feature.

[0036] In one embodiment, step 104 includes: Step 1041: Add the current weight of each source information in the weighted fusion feature to the element value at the corresponding position in the dynamic contribution vector to obtain the adjusted weight, wherein each element in the dynamic contribution vector corresponds to the weight adjustment value of each source information in the weighted fusion feature. First, obtain the dynamic contribution vector generated in step 1036, and the current weights of each source information in the weighted fusion feature. Each element in the dynamic contribution vector corresponds to the weight adjustment value of each source information in the weighted fusion feature. That is, each source information has a corresponding dynamic contribution value, which quantifies the magnitude of adjustment that the source information should be under the current abnormal fluctuation environment. Then, add the current weight of each source information to the element value at the corresponding position in the dynamic contribution vector to obtain the adjusted weight (for example, if the current weight of wind speed source information is 0.5, and the element value at the corresponding position in the dynamic contribution vector is +0.2, then the adjusted weight after adding is 0.7, indicating that the importance of wind speed information in the current environment has been increased; while the current weight of temperature source information is 0.3, and the element value at the corresponding position in the dynamic contribution vector is -0.1, then the adjusted weight after adding is 0.2, indicating that the importance of temperature information in the current environment has been reduced). Therefore, the weight adjustment values ​​in the dynamic contribution vector are directly added to the current weights of each source information through addition operations, realizing the dynamic correction of the weights of each source information. This allows the corrected weights to reflect the true contribution of each source information to the prediction task under the current environmental fluctuations in real time, providing an accurate weight basis for subsequent reweighting and fusion.

[0037] Step 1042: Re-weight and fuse the multi-source information using the adjusted weights to obtain the weighted fusion features after weight adjustment; First, obtain the adjusted weights corresponding to each source information obtained in step 1041, and the original data sequences corresponding to each source information in the multi-source information collected in step 1011. The adjusted weights have incorporated the correction information of the dynamic contribution vector to the current environmental fluctuations, reflecting the true importance of each source information at the current moment. Then, use the adjusted weights to re-weight and fuse the multi-source information. That is, for each source information, multiply each value in its original data sequence by the adjusted weight corresponding to that source information, and then sum the weighted results of all source information item by item (for example, the adjusted weight of wind speed source information is 0.7, and its collected wind speed value is 15.0 m / s, so the weighted contribution is 10.5; the adjusted weight of temperature source information is 0.2, and its collected temperature value is 25.5℃, so the weighted contribution is 5.1, and the sum of the two is 15.6). After completing the above calculations, the weighted fusion feature with adjusted weights is obtained. Therefore, the adjusted weights are actually applied to the reweighted fusion process of multi-source information, so that the weighted fusion features after weight adjustment can accurately reflect the dynamic importance distribution of information from each source under abnormal fluctuation environment, and provide high-quality input with real-time weight correction for the subsequent secondary training of the long short-term memory network.

[0038] Step 1043: Use the weighted fusion features after weight adjustment as training samples and input them into the Long Short-Term Memory network. Update the network parameters of the Long Short-Term Memory network through the backpropagation algorithm. First, the weighted fusion feature obtained in step 1042 with adjusted weights is acquired. This feature has been dynamically corrected in real time to adjust the weights of each source of information using a dynamic contribution vector, and can reflect the true importance distribution of each source of information under the current abnormal fluctuation environment. This feature is organized into training samples for model parameter optimization. Then, the training samples are input into the Long Short-Term Memory network, and a complete forward propagation calculation is performed to obtain the prediction output under the current network parameters. Then, with the error between the prediction output and the true value as the target, the gradient of each network parameter is calculated through the backpropagation algorithm, and the network parameters of the Long Short-Term Memory network (including the weight matrix and bias terms, where the internal weights of the input gate, forget gate, and output gate are updated synchronously) are updated along the gradient descent direction. For example, in a strong wind disturbance scenario, the weight of the corrected wind speed information is increased, and it is used as a training sample input into the network. The backpropagation algorithm adjusts the dimension values ​​related to the wind speed feature in the internal weight matrix of the network accordingly based on the contribution of the wind speed information to the prediction error. Therefore, the backpropagation algorithm transmits the dynamic weight information contained in the corrected fusion features to the various gating mechanisms inside the Long Short-Term Memory network, so that the network parameters of the Long Short-Term Memory network are adjusted to adapt to the current abnormal fluctuation environment, laying the parameter foundation for the convergence of subsequent secondary training and the generation of optimized prediction models.

[0039] Step 1044: Calculate the loss function value of the Long Short-Term Memory network. If the loss function value meets the preset convergence condition, stop training and output the current network parameters as the optimized prediction model. If the loss function value does not meet the preset convergence condition, continue to perform backpropagation to update the network parameters and repeat training until the preset convergence condition is met.

[0040] First, after completing a network parameter update in step 1043, the loss function value of the current Long Short-Term Memory network on the validation set is calculated (for example, using mean squared error as the loss function, the mean squared error between the current predicted output and the true value is calculated, resulting in a loss function value of 0.035); then, it is determined whether the loss function value meets a preset convergence condition (for example, the preset convergence condition is that the loss function value is less than 0.05 or the decrease in the loss function value is less than 0.001 for 5 consecutive iterations); if the loss function value meets the preset convergence condition (such as the loss function...), the network is considered to have a loss function value of 0.035. If the loss function value is 0.035 (less than 0.05), training stops, the current network parameters of the Long Short-Term Memory network are fixed and output as the optimized prediction model, indicating that the current model has fully learned the dynamic patterns in the corrected fused features. If the loss function value does not meet the preset convergence condition (e.g., the loss function value is 0.06 but still greater than 0.05), backpropagation is continued to update the network parameters, and the updated parameters are used for the next round of forward propagation calculation. Training is repeated (i.e., the process of forward propagation, loss calculation, backpropagation, and parameter update is repeated) until the preset convergence condition is met. Thus, the training process is closed-loop controlled by the preset convergence condition, ensuring that the secondary training automatically terminates after the model performance reaches the expected standard, avoiding prediction bias caused by undertraining or overfitting caused by overtraining, and achieving a balance between prediction accuracy and generalization ability in the output optimized prediction model.

[0041] In one embodiment, step 105 includes: Step 1051: Input the real-time collected multi-source information into the optimized prediction model, and output the prediction result after forward calculation of the model; First, real-time multi-source information is acquired (e.g., wind speed from a wind speed sensor in the distribution network is 15.0 m / s, temperature from a temperature sensor is 25.5℃, and atmospheric pressure from a weather station is 1013 hPa). This real-time data is organized into a multi-source information sequence in the same format as in step 1011, and used as input to the optimized prediction model. Then, the real-time multi-source information sequence is input into the optimized prediction model, and forward computation is performed within the model. That is, the data passes through the gating mechanisms and network layers of the pre-trained Long Short-Term Memory network sequentially from the input layer, and is calculated layer by layer, finally outputting the prediction result (e.g., the model outputs the probability distribution of the distribution network operating state at a future time step, where the probability of normal state is 0.85, the probability of warning state is 0.10, and the probability of fault state is 0.05). Thus, the optimized prediction model performs rapid inference on the real-time multi-source information, mapping the current multi-source data into a quantitative prediction of the future situation, providing a data foundation for subsequent determination of environmental change trends and adjustment of state probabilities.

[0042] Step 1052: Perform time series analysis on the prediction results, calculate the rate of change index of each environmental parameter, compare the rate of change index with the preset trend threshold, and determine the current environmental change trend. First, the prediction results output in step 1051 are obtained. Time series analysis is performed on the time series components corresponding to each environmental parameter in the prediction results. Specifically, the rate of change of each environmental parameter at the current moment relative to multiple previous historical moments is calculated to obtain the rate of change index corresponding to each environmental parameter (for example, based on the wind speed prediction value sequence continuously collected in the past 5 minutes, the rate of change of the current wind speed relative to 5 minutes ago is calculated to be +2.5 m / s², while the rate of change of temperature is -0.3℃ / min). Then, the rate of change index of each environmental parameter is compared with a preset trend threshold. When the positive or negative absolute value of the rate of change index exceeds the corresponding threshold, the current environmental change trend is determined to be an upward trend, a downward trend, or a violent fluctuation trend (for example, if the wind speed change rate of +2.5 m / s² exceeds the preset upward trend threshold of +1.5 m / s², the current environmental change trend is determined to be "rapid upward trend of wind speed"). Therefore, by calculating the rate of change of environmental parameter components in the prediction results and comparing it with the threshold, the static prediction value output by the model is transformed into a dynamic judgment of the direction of change of the surrounding environment of the power distribution network, providing a basis for trend judgment for the subsequent construction of the state transition matrix and the adjustment of the state probability.

[0043] Step 1053: Construct a state transition matrix based on the environmental change trend, and calculate the adjusted state probability distribution based on the state transition matrix using the Markov chain method; First, based on the environmental change trend determined in step 1052 (e.g., a "rapid increase in wind speed"), the transition characteristics of the distribution network operating state under this trend are determined, and a state transition matrix is ​​constructed. The rows and columns of the state transition matrix correspond to a preset state space (e.g., defining the distribution network operating state as three discrete states: "normal state," "warning state," and "fault state"). Each element in the matrix represents the probability value of transitioning from one state to another under a given environmental change trend (e.g., under a "rapid increase in wind speed" trend, the probability of transitioning from "normal state" to "warning state"). The probability of transitioning from "warning state" to "fault state" is 0.3, with a probability of 0.6. Then, based on the state transition matrix, a Markov chain method is used to iteratively calculate the current state probability. This involves multiplying the probability distribution of each state at the current moment by the state transition matrix to calculate the adjusted state probability distribution (for example, if the current probability of normal state is 0.7, warning state is 0.2, and fault state is 0.1, after one iteration of the Markov chain, the adjusted state probability distribution is 0.5 for normal state, 0.35 for warning state, and 0.15 for fault state). Thus, the environmental change trend is quantified into a state transition matrix, and the Markov chain method transmits this trend information to the state probability update process, achieving a quantitative estimate of the operating state probability under dynamic environmental changes. This provides a quantifiable probability benchmark for subsequent comparison with historical records.

[0044] Step 1054: Compare the adjusted state probability distribution with the state probability vectors at each time point in the historical record to calculate the matching degree, and assign attention weights to each historical record based on the matching degree. First, obtain the adjusted state probability distribution calculated in step 1053 (e.g., the adjusted state probability distribution is 0.5 for normal state, 0.35 for warning state, and 0.15 for fault state). Extract the state probability vectors stored at each time point from the historical records (each historical record contains the probability distribution of each state at the corresponding time). Compare the adjusted state probability distribution with the state probability vectors of each historical record to calculate the matching degree between the adjusted state probability distribution and each historical state probability vector. Specifically, Euclidean distance or cosine similarity can be used as the quantitative index of matching degree (e.g., if the state probability vector in a certain historical record is 0.55 for normal, 0.30 for warning, and 0.15 for fault, and the Euclidean distance between the current adjusted state probability distribution and the historical record is small, then the matching degree is high). Then, based on the matching degree corresponding to each historical record, use the softmax function to normalize each matching degree and assign attention weights to each historical record. The historical record with the higher matching degree receives a higher attention weight (e.g., the historical record with the highest matching degree receives an attention weight of 0.5, the second highest receives 0.3, and the remaining records share the remaining weights). Therefore, the similarity between the current state distribution and historical records is quantified by matching degree calculation, and attention weight is used to make the historical records with high similarity play a dominant role in the subsequent extraction of related paths, so as to ensure that the historical experience most relevant to the current situation is given priority, thereby improving the correlation between the situation prediction results and historical experience.

[0045] Step 1055: Extract relevant state sequences from historical records based on the attention weights to construct focus-related paths; perform path dependency modeling on the focus-related paths, extract the state transition relationships of each associated node in the path, and dynamically correct the state transition relationships by combining real-time collected multi-source information to obtain situation prediction results.

[0046] First, based on the attention weights assigned in step 1054, historical records with high attention weights are selected from the historical records. Relevant state sequences are extracted from these historical records (e.g., high-weight historical records document the complete state evolution process of the power distribution network from a normal state through an early warning state to a fault state under similar rapid wind speed increases). These relevant state sequences are organized into focal correlation paths in chronological order. Then, path dependency modeling is performed on these focal correlation paths, i.e., analyzing the dependence strength and transition patterns between adjacent states in the path, extracting the state transition relationships of each associated node in the path (e.g., the transition probability from normal to early warning is 0.7, and the transition probability from early warning to fault is 0.4 from the focal correlation paths). The state transition relationships are then dynamically corrected using real-time multi-source information (e.g., real-time monitoring data showing that the current wind speed is still rising). (e.g., if real-time information shows that the current wind speed increase rate is higher than the historical average for the same period, the transition probability from early warning to fault is increased from 0.4 to 0.55). The corrected state transition relationships are applied to the iterative update of the current state probability, ultimately yielding the situation prediction result. Therefore, the focus-related path locks in the historical evolution pattern most relevant to the current situation through attention weights, while the combination of path dependency modeling and real-time multi-source information correction enables historical patterns to be adaptively adjusted under the current environmental changes. The output situation prediction results take into account the reference value of historical experience and the dynamic changes of the real-time environment, thus improving the accuracy and environmental adaptability of the prediction.

[0047] This invention provides a method for predicting the status of power distribution network dispatching operations. Addressing the challenges of multi-source heterogeneous data fusion, delayed response to abnormal fluctuations, and insufficient adaptability of prediction models to dynamic environments in power distribution network dispatching operations, the method first collects multi-source information transmitted from sensors and weather stations in the power distribution network and assigns initial weights to each source using an attention mechanism. Weighted fusion of the multi-source information yields a weighted fusion feature, allowing the initial weights to be dynamically quantified based on the correlation between each source and the prediction target, avoiding the omission or interference of key information caused by fixed weight allocation. Next, the weighted fusion feature is input into a Long Short-Term Memory (LSTM) network, and its gating mechanism extracts the temporal correlation information at each time step to obtain a hidden state sequence. This allows the network to selectively retain or discard information, achieving accurate representation of temporal dependencies. Finally, when a hidden state sequence value deviates from a preset range, a sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The system dynamically updates the attention weight matrix based on local fluctuations to generate a dynamic contribution vector, enabling the attention weight matrix to adaptively update according to the degree of environmental fluctuations, thus solving the problem of static weights lacking dynamic adaptability. Then, the weights of each source of information are adjusted according to the dynamic contribution vector, and the weighted fusion features with adjusted weights are input again into the Long Short-Term Memory (LSTM) network for secondary training to update the network parameters, resulting in an optimized prediction model. This optimizes the model parameters to adapt to the current abnormal fluctuation environment, solving the problem of delayed response to environmental changes caused by fixed model parameters. Finally, real-time multi-source information is input into the optimized prediction model to obtain prediction results. Based on the prediction results, the environmental change trend is determined to adjust the state probability distribution. Based on the adjusted state probability distribution and historical records, key correlation paths are strengthened through an attention mechanism to output situational prediction results, ensuring that the prediction output takes into account both the reference value of historical experience and the dynamic changes of the real-time environment. The entire solution, through the synergistic mechanism of dynamic weight adjustment and model parameter optimization, solves the technical problems of insufficient prediction accuracy and scheduling reliability caused by static weights in multi-source information fusion and fixed model parameters, significantly improving the situational awareness accuracy and anomaly response speed of distribution network scheduling in complex and ever-changing environments. In summary, this invention solves the problem that existing technologies use static weights and fixed model parameters in multi-source information fusion, which cannot dynamically adapt to environmental changes, resulting in poor prediction accuracy and scheduling reliability.

[0048] The above is a method for predicting the status of power distribution network dispatching operations provided in the embodiments of the present invention. The following is a system for predicting the status of power distribution network dispatching operations provided in the embodiments of the present invention.

[0049] Please see Figure 2 The present invention provides a distribution network dispatching operation status prediction system, comprising: Construction unit 201 is used to collect multi-source information transmitted by sensors and weather stations in the power distribution network, assign initial weights to each source information using an attention mechanism and construct an attention weight matrix, and use the attention weight matrix to perform weighted fusion of the multi-source information to obtain weighted fusion features; Extraction unit 202 is used to input the weighted fusion features into a long short-term memory network, and extract the temporal correlation information of each time step through the gating mechanism of the long short-term memory network to obtain the hidden state sequence; The generation unit 203 is used to perform anomaly detection on the hidden state sequence. When the sequence value is detected to deviate from the preset range, a sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The attention weight matrix is ​​updated according to the local fluctuation amplitude to generate a dynamic contribution vector. Training unit 204 is used to adjust the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and input the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters and obtain an optimized prediction model. The output unit 205 is used to input the multi-source information collected in real time into the optimized prediction model, obtain the prediction result, determine the environmental change trend based on the prediction result, adjust the state probability distribution based on the environmental change trend, strengthen the key correlation path through the attention mechanism based on the adjusted state probability distribution and historical records, and output the situation prediction result.

[0050] Furthermore, this embodiment of the invention also provides a distribution network dispatching operation status prediction device, the device including a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the steps of the power distribution network dispatching operation status prediction method as described in the above method embodiments, according to the instructions in the program code.

[0051] In one embodiment, the device can be a standalone distribution network dispatching operation status prediction terminal device, or a new functional module integrated into an existing distribution network dispatching system. The memory includes, but is not limited to, non-transitory computer-readable storage media such as read-only memory, random access memory, magnetic disks, or optical disks, used to store pre-written program code and intermediate data generated during operation (such as weighted fusion features, hidden state sequences, dynamic contribution vectors, optimized prediction model parameters, etc.), and to transmit the program code to the processor. The processor can be a central processing unit (CPU), graphics processing unit (GPU), application-specific integrated circuit (ASIC), or field-programmable gate array (FPGA), used to load and execute the program code stored in the memory. In actual operation, when situation prediction is required, the processor reads the program code from memory and performs the following operations sequentially: It collects multi-source information from sensors and weather stations deployed in the distribution network; it performs initial weight allocation and weighted fusion using an attention mechanism; it calls a long short-term memory network to extract temporal correlations to obtain hidden state sequences; it performs anomaly detection and updates the attention weight matrix based on sliding window analysis to generate a dynamic contribution vector; it adjusts the weights based on the dynamic contribution vector and retrains the long short-term memory network to obtain an optimized prediction model; finally, it outputs predictions based on the optimized prediction model for the real-time collected multi-source information. Through the above equipment, the distribution network scheduling operation situation prediction method can be implemented in hardware, ensuring efficient execution of the prediction process and accurate output of results.

[0052] Furthermore, this embodiment of the invention also provides a computer-readable storage medium for storing program code, which is used to execute the power distribution network dispatching operation status prediction method described in the above method embodiments.

[0053] In one embodiment, the computer-readable storage medium includes, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. The program code is stored in the computer-readable storage medium in the form of computer-executable instructions. When the program code is loaded and executed by the processor of the computer device, the processor performs the following operations: Collecting multi-source information transmitted by sensors and weather stations in the power distribution network; allocating initial weights and constructing an attention weight matrix using an attention mechanism; weighted fusing of the multi-source information to obtain weighted fusion features; inputting the weighted fusion features into a long short-term memory network; extracting temporal correlation information through a gating mechanism to obtain a hidden state sequence; performing anomaly detection on the hidden state sequence; initiating sliding window analysis when a sequence value deviates from a preset range; extracting the local fluctuation amplitude of the sequence within the window and updating the attention weight matrix to generate a dynamic contribution vector; adjusting the weights of each source information according to the dynamic contribution vector; inputting the weighted fusion features again into the long short-term memory network for secondary training to update network parameters and obtain an optimized prediction model; inputting the real-time collected multi-source information into the optimized prediction model to obtain prediction results; determining the environmental change trend and adjusting the state probability distribution in conjunction with historical records; strengthening key correlation paths through an attention mechanism; and outputting the situation prediction results. In practical applications, this computer-readable storage medium can be sold or distributed as a standalone software product. After purchase, users can install the program code from the storage medium into their existing power distribution network dispatching system to achieve the aforementioned functions without replacing hardware, thus reducing the deployment cost of the technical solution. When the program code is executed by a processor using this storage medium, it can fully implement each step of the power distribution network dispatching operation situation prediction method, achieving the technical effects of improving prediction accuracy and environmental adaptability.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0056] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0058] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the operational status of a distribution network dispatching system, characterized in that, include: Multi-source information transmitted by sensors and weather stations in the power distribution network is collected. An attention mechanism is used to assign initial weights to each source of information and construct an attention weight matrix. The attention weight matrix is ​​then used to perform weighted fusion of the multi-source information to obtain weighted fusion features. The weighted fusion features are input into a long short-term memory network, and the temporal correlation information of each time step is extracted through the gating mechanism of the long short-term memory network to obtain the hidden state sequence. Anomaly detection is performed on the hidden state sequence. When the sequence value deviates from the preset range, sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The attention weight matrix is ​​updated based on the local fluctuation amplitude to generate a dynamic contribution vector. The weights of each source information in the weighted fusion feature are adjusted according to the dynamic contribution vector. The weighted fusion feature with adjusted weights is then input into the long short-term memory network for secondary training to update the network parameters and obtain the optimized prediction model. The multi-source information collected in real time is input into the optimized prediction model to obtain the prediction results. The environmental change trend is determined based on the prediction results. The state probability distribution is adjusted according to the environmental change trend. Based on the adjusted state probability distribution and historical records, the key correlation path is strengthened through an attention mechanism, and the situation prediction result is output.

2. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The method involves collecting multi-source information transmitted from sensors and weather stations in the power distribution network, assigning initial weights to each source of information using an attention mechanism, constructing an attention weight matrix, and then using the attention weight matrix to perform weighted fusion of the multi-source information to obtain weighted fusion features, including: Environmental data is collected by sensors deployed in the power distribution network, and real-time meteorological indicators are obtained from meteorological stations to form a multi-source information sequence; The multi-source information sequence is preprocessed to convert the information from each source into a feature vector in numerical form; A query vector is generated based on the current prediction target. The feature vectors of each source information are input into the attention mechanism to generate corresponding key vectors. The dot product similarity between the query vector and the key vectors corresponding to each source information is calculated. After scaling, the similarity is normalized by the softmax function to obtain the attention score of each source information as the initial weight. An attention weight matrix is ​​constructed based on the initial weights of each source information. The feature vectors of each source information are then weighted and summed using the attention weight matrix to obtain the weighted fusion features.

3. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The weighted fusion features are input into a Long Short-Term Memory (LSTM) network, and the temporal correlation information of each time step is extracted through the gating mechanism of the LSM network to obtain the hidden state sequence, including: The weighted fusion features are arranged according to time steps to construct a time-series input sequence; The time-series input sequence is sequentially input into the Long Short-Term Memory (LSTM) network. At each time step, the input gate controls the amount of new information written, the forget gate controls the amount of cell state retained from the previous time step, and the output gate controls the amount of hidden state output at the current time step, thus updating the cell state and hidden state. The hidden state output at each time step is recorded to form a hidden state sequence.

4. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The anomaly detection of the hidden state sequence involves initiating a sliding window analysis when a sequence value deviates from a preset range, extracting the local fluctuation amplitude of the sequence within the window, including: Obtain the hidden state value at each time step in the hidden state sequence, and calculate the difference in hidden state values ​​between adjacent time steps; The difference is compared with a preset threshold. When the difference exceeds the preset threshold, it is determined that the current sequence value deviates from the preset range and a sliding window analysis is started. The window length of the sliding window is determined, and the window is slid on the hidden state sequence with a fixed step size to extract the hidden state subsequence within the current window. Calculate the mean and variance of each hidden state value in the hidden state subsequence, and determine the local fluctuation amplitude within the current window based on the mean and variance.

5. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The step of updating the attention weight matrix based on the local fluctuation amplitude to generate a dynamic contribution vector includes: The weight adjustment factor corresponding to each source of information is calculated based on the local fluctuation amplitude within the current window, and the weight adjustment factor is positively correlated with the local fluctuation amplitude; The weight adjustment factor corresponding to each source information is multiplied by the initial weight at the corresponding position in the attention weight matrix to obtain the updated attention weight matrix; The updated attention weight matrix is ​​weighted and calculated with the feature vectors of each source information in the weighted fusion feature to generate a dynamic contribution vector.

6. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The step of adjusting the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and inputting the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters, to obtain an optimized prediction model, includes: The current weight of each source information in the weighted fusion feature is added to the element value at the corresponding position in the dynamic contribution vector to obtain the adjusted weight, wherein each element in the dynamic contribution vector corresponds to the weight adjustment value of each source information in the weighted fusion feature. The multi-source information is re-weighted and fused using the adjusted weights to obtain the weighted fusion features with adjusted weights. The weighted fusion features after weight adjustment are used as training samples and input into the long short-term memory network. The network parameters of the long short-term memory network are updated through the backpropagation algorithm. Calculate the loss function value of the Long Short-Term Memory network. If the loss function value satisfies the preset convergence condition, stop training and output the current network parameters as the optimized prediction model. If the loss function value does not satisfy the preset convergence condition, continue to perform backpropagation to update the network parameters and repeat training until the preset convergence condition is satisfied.

7. The method for predicting the status of power distribution network dispatching operations according to claim 1, characterized in that, The process involves inputting real-time collected multi-source information into the optimized prediction model to obtain prediction results, determining environmental change trends based on the prediction results, adjusting the state probability distribution based on the environmental change trends, and strengthening key correlation paths through an attention mechanism based on the adjusted state probability distribution and historical records to output situation prediction results, including: The multi-source information collected in real time is input into the optimized prediction model, and the prediction result is output after forward calculation of the model. The prediction results are subjected to time series analysis to calculate the rate of change of each environmental parameter. The rate of change is then compared with a preset trend threshold to determine the current environmental change trend. A state transition matrix is ​​constructed based on the environmental change trend, and the adjusted state probability distribution is calculated using the Markov chain method based on the state transition matrix. The adjusted state probability distribution is compared with the state probability vectors at each time point in the historical record to calculate the matching degree, and attention weights are assigned to each historical record based on the matching degree. Based on the attention weights, relevant state sequences are extracted from historical records to construct focus-related paths; path dependency modeling is performed on the focus-related paths to extract the state transition relationships of each associated node in the path, and the state transition relationships are dynamically corrected by combining real-time multi-source information to obtain situation prediction results.

8. A distribution network dispatching operation status prediction system, characterized in that, include: The construction unit is used to collect multi-source information transmitted by sensors and weather stations in the power distribution network, and uses an attention mechanism to assign initial weights to each source of information and construct an attention weight matrix. The attention weight matrix is ​​then used to perform weighted fusion of the multi-source information to obtain weighted fusion features. The extraction unit is used to input the weighted fusion features into the long short-term memory network, and extract the temporal correlation information of each time step through the gating mechanism of the long short-term memory network to obtain the hidden state sequence; The generation unit is used to perform anomaly detection on the hidden state sequence. When the sequence value is detected to deviate from the preset range, a sliding window analysis is initiated to extract the local fluctuation amplitude of the sequence within the window. The attention weight matrix is ​​updated according to the local fluctuation amplitude to generate a dynamic contribution vector. The training unit is used to adjust the weights of each source information in the weighted fusion feature according to the dynamic contribution vector, and input the weighted fusion feature with adjusted weights into the long short-term memory network for secondary training to update the network parameters and obtain the optimized prediction model. The output unit is used to input multi-source information collected in real time into the optimized prediction model, obtain prediction results, determine the environmental change trend based on the prediction results, adjust the state probability distribution based on the environmental change trend, strengthen key correlation paths through an attention mechanism based on the adjusted state probability distribution and historical records, and output the situation prediction result.

9. A distribution network dispatching operation status prediction device, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the power distribution network dispatching operation status prediction method according to any one of claims 1-7 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code, which is used to execute the distribution network dispatching operation status prediction method according to any one of claims 1-7.

Images