A method for calculating the capacity-to-load ratio of medium-voltage distribution networks based on deep neural networks

Abstract

This invention proposes a method for calculating the capacity-to-load ratio of medium-voltage distribution networks based on deep neural networks. The method includes: collecting multi-source data from the medium-voltage distribution network; designing an adaptive hierarchical clustering algorithm to dynamically identify load types based on a high-dimensional feature matrix; constructing a time-series multi-level recurrent neural network for load prediction; calculating the current capacity-to-load ratio of the network based on the load prediction and performing preliminary optimization; constructing a deep learning model based on the preliminary optimized load prediction to further optimize energy management strategies; and monitoring the operation of the medium-voltage distribution network in real time, using a reinforcement learning model to optimize the deep learning model and energy management strategies. This invention can significantly improve the accuracy and real-time performance of calculating the capacity-to-load ratio of medium-voltage distribution networks, overcome the shortcomings of existing technologies in dynamically adjusting and optimizing under complex load environments, and effectively ensure the stable operation of the power grid and power supply security.

Description

Technical Field

[0001] This invention belongs to the technical field of deep neural networks, and particularly relates to a method for calculating the capacity-to-load ratio of medium-voltage distribution networks based on deep neural networks. Background Technology

[0002] Medium-voltage distribution networks are a crucial component of the power system, responsible for transmitting electrical energy from substations to various user terminals. In modern urban power supply systems, the load variations of medium-voltage distribution networks are complex and diverse, influenced by factors such as seasonality, weather, and user behavior, resulting in a high degree of uncertainty in their operational status. To ensure power supply security and power quality, distribution network operators need to accurately determine the grid's capacity utilization ratio (CUP), which is the ratio between the actual load capacity of transformers and lines and their rated capacity. Precise calculation and dynamic adjustment of the CUP not only contribute to the rational allocation of power resources but also effectively prevent equipment overload and power outages, improving the reliability and economy of the power grid.

[0003] Currently, traditional methods for calculating the capacity-to-load ratio in medium-voltage distribution networks mainly rely on static models and fixed parameter settings. These methods are typically based on historical data and empirical rules, and cannot respond to dynamic load changes in real time. With the widespread adoption of distributed energy sources, smart meters, and various new types of power equipment in distribution networks, the load patterns of the power grid have become more diverse and complex, making traditional static models unable to effectively adapt to rapidly changing power demands. Furthermore, traditional methods lack comprehensive consideration of factors such as grid topology, load type, and time-series changes during the calculation process, resulting in low accuracy in capacity-to-load ratio prediction and difficulty in dealing with the complex nonlinear dynamic behavior of power systems.

[0004] Another shortcoming of existing technologies lies in the lack of responsiveness to real-time load fluctuations and the absence of intelligent adjustment mechanisms. Traditional capacity-to-load ratio calculation methods cannot fully utilize the increasingly abundant real-time data and environmental information in the distribution network, making it difficult to accurately predict and optimize scheduling under different load conditions. Due to changes in the power grid operating environment and the volatility of electricity demand, these methods are slow to react and untimely in response to sudden load changes and equipment failures, which may lead to local overloads or uneven load distribution in the power grid, thereby triggering a chain reaction and affecting the stability and security of the entire power grid. Summary of the Invention

[0005] The purpose of this invention is to design a method for calculating the capacity-to-load ratio of medium-voltage distribution networks based on deep neural networks. By combining the latest time-series deep learning technology and load type recognition technology, this invention can predict the load changes of the power grid in real time and perform dynamic optimization scheduling of the load through an intelligent energy management system. This method overcomes the limitations of traditional capacity-to-load ratio calculation models and can provide higher prediction accuracy and optimization capabilities in complex and ever-changing power environments, significantly improving the operating efficiency and security of the power grid.

[0006] To achieve the above objectives, this invention provides a method for calculating the capacity-to-load ratio of medium-voltage distribution networks based on deep neural networks. The method includes the following steps:

[0007] S1. Collect multi-source data from the medium-voltage distribution network and construct the first high-dimensional feature matrix based on the multi-source data;

[0008] S2. Design an adaptive hierarchical clustering algorithm to dynamically identify load types based on a high-dimensional feature matrix. Then, extract the time-series features of the load and the autocorrelation features from the time-series features from the load types, and construct a second high-dimensional feature matrix from the time-series features and the autocorrelation features.

[0009] S3. Construct a time-series multi-level recurrent neural network, using the second high-dimensional feature matrix as input, and perform load prediction by combining short-term and long-term features through recursive structure and regularization methods.

[0010] The temporal multi-level recurrent neural network includes a first recurrent network and a second recurrent network.

[0011] The first layer of the recursive network processes the temporal characteristics of each load type, as shown below:

[0012] h1(t)=RNN1(T C (t))+λ1·reg(T C (t))

[0013] Where h1(t) is the hidden state of the first recurrent network, λ1 is the regularization coefficient, and reg(T) is the hidden state of the first recurrent network. C (t) is a specially designed local smoothing regularization term, RNN1 is the first layer of the recursive network, and T C (t) represents the temporal feature; where the local smoothing regularization term reg(T) is... C (t) is represented as follows:

[0014]

[0015] Where n is the total number of windows, i is the window size, and t is the current time.

[0016] The second recurrent network combines the hidden states and autocorrelation features of the first layer to process the long-term trend of the load, as shown below:

[0017] h2(t)=RNN2(h1(t),A C (t))+λ2·reg(A C (t))

[0018] Where h2(t) is the hidden state of the second recursive network, A C (t) is the autocorrelation characteristic, and λ² is the regularization coefficient; reg(A C (t) is the long-term periodic regularity term, expressed as follows:

[0019]

[0020] Where m is the total number of time steps, and k is the k time steps;

[0021] After passing through the output of the aforementioned temporal multi-level recurrent neural network, the final hidden state h2(t) is passed to a fully connected layer for mapping, generating the load prediction value P(t), as follows:

[0022] P(t) = W p ·h2(t)+b p

[0023] Among them, W p It is the weight matrix of the fully connected layer, b p It is the bias term, where P(t) represents the predicted load value at future time step t;

[0024] S4. Calculate the current capacity ratio of the power grid based on the load forecast, and perform preliminary optimization on the current capacity ratio to obtain a preliminary optimized load forecast.

[0025] Specifically, the preliminary optimization of the current capacity ratio to obtain a preliminary optimized load prediction includes:

[0026] Construct an adaptive load balancing optimization model based on the capacity ratio R of each node. i (t), load balancing is performed on nodes whose load exceeds a set threshold, as shown below:

[0027]

[0028] Where, ΔP i(t) is the node load adjustment term, which is dynamically adjusted based on the current capacity ratio of each node and the load status of its neighboring nodes; α is the load balancing adjustment coefficient, used to control the load distribution speed; β is the neighboring node adjustment coefficient, used to adjust the load difference with neighboring nodes; N(i) is the set of nodes adjacent to node i, and |N(i)| is the number of neighboring nodes; P h (t) is the predicted load value of the neighboring node j; Capacity i This refers to the capacity of the power grid node corresponding to this load type; P i (t) is the predicted value of the i-th type of load;

[0029] At the same time, a dynamic adjustment function D(t) for the power grid is introduced to describe the dynamic rate of change of the current capacity-to-load ratio in the power grid, as follows:

[0030]

[0031] Based on the current value of the power grid dynamic adjustment function D(t), identify the areas of load change in the power grid and perform load scheduling in advance according to the degree of change;

[0032] S5. Based on the preliminary optimized load prediction, a deep learning model is constructed to further optimize the energy management strategy, and the model parameters of the deep learning model are adjusted to adapt to the actual operating conditions.

[0033] S6. Monitor the operation of the medium-voltage distribution network in real time, and use reinforcement learning models to optimize deep learning models and energy management strategies.

[0034] Furthermore, the multi-source data includes electrical data, environmental data, and user behavior data;

[0035] The electrical data includes voltage, current, and power;

[0036] The environmental data refers to environmental changes.

[0037] The user behavior data refers to load fluctuations caused by different users' electricity consumption patterns. These load fluctuations include load power factor and electricity consumption characteristics.

[0038] The construction of the first high-dimensional feature matrix based on multi-source data specifically includes:

[0039] The multi-source data is preprocessed, including data synchronization and normalization, noise filtering and outlier handling, and feature expansion and time-series feature generation. Finally, time-series features are generated and integrated into a first high-dimensional feature matrix.

[0040] Furthermore, the adaptive hierarchical clustering algorithm is designed to dynamically identify load types based on a high-dimensional feature matrix, specifically including:

[0041] The first high-dimensional feature matrix is ​​segmented according to a fixed window to form a segmented matrix S(t) containing the past n time steps. Then, principal component analysis is used to reduce the dimensionality of the segmented matrix S(t) to generate feature vector R(t). Finally, an adaptive hierarchical clustering algorithm is introduced to perform preliminary clustering of feature vector R(t) through distance metric and adjust the clustering level according to the time series characteristics. Finally, the load type C(t) is output.

[0042] The step of extracting the time-series features of the load and the autocorrelation features from the time-series features from the load type, and constructing a second high-dimensional feature matrix from the time-series features and the autocorrelation features, specifically includes:

[0043] For each load type C(t), the sliding window method is used to extend its characteristics, forming the time series characteristics T for each load type. C (t), and simultaneously calculate the autocorrelation function A under load type C(t). C (t), identifying the periodic characteristics of long-term loads, and ultimately using the autocorrelation function A C (t) and the timing characteristics T for each load type C (t) is reassembled into the second high-dimensional feature matrix E(t).

[0044] Furthermore, in S3, a load allocation constraint term is added to the prediction output layer of the time-series multi-level recurrent neural network to constrain the load to meet the grid capacity, as shown below:

[0045]

[0046] Where C(t) is a constraint term, P i (t) is the predicted value of the i-th type of load, Capacity i It is the corresponding distribution network capacity.

[0047] Furthermore, during the training process of the aforementioned temporal multi-level recurrent neural network, supervised learning is performed using historical load data, and a weighted mean square error loss function is employed to assign different weights to errors of different load types, as shown below:

[0048]

[0049] Where Y(t) is the actual load, P(t) is the predicted load, and α is a weighting factor that is dynamically adjusted according to the load type.

[0050] Furthermore, the capacity ratio R of each node i (t) is the ratio between the load of each node in the power grid and its available capacity, used to measure whether the power grid is under normal load conditions, and is expressed as follows:

[0051]

[0052] Among them, P i (t) is the predicted value of the i-th type of load; Capacity i R(t) is the capacity of the grid node corresponding to this load type; R(t) is the total capacity-to-load ratio at time t.

[0053] Furthermore, S5 specifically includes:

[0054] S5.1 An adaptive load optimization mechanism is designed to perform real-time correction of the capacity-to-load ratio of each node, ensuring that the capacity-to-load ratio of each node is within its capacity range. Specifically, the adaptive load optimization mechanism ensures load balance by minimizing the adaptive optimization function O(t) based on the capacity-to-load ratio, while dynamically adjusting the load forecast value to make the actual operating state of the power grid as close as possible to the load forecast result. The adaptive optimization function O(t) is expressed as follows:

[0055]

[0056] Among them, R i (t) is the load capacity ratio of the i-th node; P i (t) is the predicted load value for this node; Capacity i This is the capacity of the node; ΔP i (t) is the node load adjustment amount; γ is the weighting coefficient of the balanced load adjustment term, which is used to control the relative importance of capacity ratio optimization and node load adjustment.

[0057] S5.2 An adaptive regression update mechanism is designed to adapt the deep learning model to changes in load.

[0058] S5.3. Implement closed-loop feedback for the deep learning model, feeding back the load prediction error and capacity ratio to the scheduling system, and dynamically adjusting the load prediction model for the next round. At the same time, feed back the node load adjustment amount for each load allocation to the deep learning model, and dynamically adjust the model parameters of the deep learning model.

[0059] Furthermore, S5.2 specifically includes:

[0060] First, calculate the current load prediction error ∈(t), as follows:

[0061]

[0062] Among them, Y i (t) is the actual load value;

[0063] The design of an update mechanism with weight decay allows deep learning models to have the best adaptability when the load fluctuates drastically, as shown below:

[0064]

[0065] Where ΔW(t) is the adjustment amount of the model weights, and η is the learning rate, used to control the speed of weight updates. It is the load ratio term, used to mitigate weight adjustments when the load approaches the node's capacity.

[0066] Furthermore, the state of the reinforcement learning model includes the capacity ratio of each node in the current power grid, the prediction error, and the load adjustment between adjacent nodes; the actions include dynamically adjusting the node load allocation and updating the model weights of the deep learning model; the reward function R(t) is expressed as follows:

[0067]

[0068] Among them, R i (t) is the load ratio of node i, with the goal of being close to 1, indicating that the load of the node is within its capacity; α is a coefficient used to adjust the smoothness of load adjustment to prevent over-adjustment from causing instability; β is used to control the impact of prediction error on reward to ensure that the model prediction accuracy continues to improve.

[0069] Furthermore, the reinforcement learning model is optimized using a deep Q-learning algorithm. It calculates the Q-value at the current time using state S(t) and action A(t), reflecting the expected cumulative reward for taking a certain action in a specific state. The Q-value update formula is as follows:

[0070]

[0071] Where η is the learning rate, used to control the update magnitude of the Q value, γ is the discount factor, which weighs the impact of current rewards on future gains, and R(t) is the current reward function, representing the operating effect under the current load scheduling.

[0072] The beneficial technical effects of the present invention are at least as follows:

[0073] (1) By introducing a time-series deep learning model (such as LSTM, GRU, or Transformer), this invention can accurately predict future changes in different load types based on historical data and real-time input. This high-precision prediction capability effectively solves the problem that traditional methods are unable to cope with the complex dynamic changes in the power system, enabling the power grid to respond to load changes in advance and prevent overload and equipment damage.

[0074] (2) This invention employs load type identification technology to subdivide the loads in the power grid into various types (such as residential, commercial, and industrial loads), and performs feature extraction and separate modeling for each type. This multi-dimensional identification and optimization method breaks through the limitations of traditional methods that rely solely on a single load feature for prediction, making the capacity ratio calculation more refined and accurate, and better adaptable to different load conditions and user behavior patterns.

[0075] (3) This invention introduces an intelligent energy management system (EMS), which can monitor the operating status and load conditions of the distribution network in real time, and perform dynamic load scheduling and optimized allocation based on the prediction results. Through the load balancing optimization algorithm, the system can quickly adjust the load distribution of each node under load changes and fault conditions, avoid grid overload and imbalance, and improve the grid's operational stability and ability to respond to emergencies.

[0076] (4) By introducing an adaptive learning mechanism, this invention can continuously optimize the deep learning model and energy management algorithm based on real-time feedback data, thereby improving the robustness and adaptability of the model. Compared with the traditional fixed parameter method, this adaptive optimization capability enables the system to continuously learn and adjust, adapting to the long-term changing trends of the power grid and new load patterns, thus improving the real-time performance and reliability of the capacity-to-load ratio calculation.

[0077] (5) This invention fully utilizes the abundant real-time data in the power distribution network, including electrical data, environmental data, and user behavior data. Through multi-source data fusion technology, this information is integrated into the model, improving the accuracy of prediction and optimization. This data-driven approach solves the problem that traditional methods cannot use real-time information for precise adjustments, making power grid management more intelligent and efficient. Attached Figure Description

[0078] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0079] Figure 1 This is a flowchart illustrating a method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network, according to an embodiment of the present invention. Detailed Implementation

[0080] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0081] In one or more embodiments, such as Figure 1 As shown, this invention discloses a method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network. The method includes steps 1 to 6, including:

[0082] S1. Collect multi-source data from the medium-voltage distribution network and construct the first high-dimensional feature matrix based on the multi-source data.

[0083] Specifically, in medium-voltage distribution networks, loads are affected by various factors; therefore, this invention requires data collection from multiple dimensions. The data includes:

[0084] Electrical data: Real-time voltage V(t), current I(t), and power P(t) are collected, which directly reflect the operating status of the power grid load.

[0085] Environmental data: Environmental changes, such as temperature T(t) and humidity H(t), will affect the power load, especially the energy consumption patterns of industrial and commercial users.

[0086] User behavior data: Different users’ electricity consumption patterns can lead to load fluctuations, so the load power factor cosφ(t) and electricity consumption characteristics L(t) are taken into consideration.

[0087] These data are collected in time-series format with timestamp t, ensuring the temporal consistency of the information. Integrating multi-source data helps to comprehensively reflect the actual operating status of the power grid, rather than being limited to the analysis of a single data source.

[0088] Furthermore, data from different sources often have different acquisition frequencies and time steps. To ensure their synchronization, a linear interpolation synchronization algorithm is used. This algorithm uses interpolation to unify data from different acquisition frequencies to the same time step Δt. The interpolation process ensures data time alignment and avoids model bias caused by data inconsistency.

[0089] After data synchronization, normalization is performed. For each data type, this invention scales it to the [0,1] interval to reduce the impact of units on the model. The normalization formula is as follows:

[0090]

[0091] Where X(t) is the original data, X min and X max These represent the minimum and maximum values ​​of the data source, respectively. After normalization, each data feature will vary at the same scale, avoiding errors during model training caused by different units of measurement. This synchronization and normalization step ensures that multi-source data can be effectively integrated, providing consistent and usable features for subsequent load forecasting.

[0092] Furthermore, power grid data is frequently subject to various interferences, especially electromagnetic interference, equipment aging, and signal jitter. To ensure data quality, an adaptive median filter is used to handle noise and outliers. This filter dynamically adjusts the filter window size to adapt to different noise levels. The core formula for the filter is:

[0093] Y(t)=median{X(ti),X(t-i+1),…,X(t+i)}

[0094] Where Y(t) is the filtered data and X(t) is the original data, the window size i is dynamically adjusted according to the noise level. Compared with traditional fixed-window filters, this method can better preserve the main features of the signal, and is particularly suitable for the fluctuating characteristics of load data. This adaptive noise filtering can not only effectively remove noise, but also preserve the key fluctuation characteristics of the power grid load, ensuring that the model training in subsequent steps is based on high-quality data.

[0095] Furthermore, load changes in the power grid exhibit significant time dependence, therefore analysis cannot rely solely on data from the current time point. To capture past load behavior, a sliding window method is used to generate time-series features. For each time step t, an expanded feature set Z(t) is created, containing data from the past n time steps:

[0096] Z(t)=[X(t),X(t-1),…,X(tn)]

[0097] Here, Z(t) represents the extended feature at time t, and the window size n can be dynamically adjusted based on the characteristics of the power grid. The optimal window size is typically selected through cross-validation. This method allows the model to learn not only the current time data but also the load variation trend over a past period. This step, by extending the time-series features, ensures that subsequent models can effectively capture the short-term and long-term fluctuations in power grid load, enhancing the accuracy of load forecasting and its ability to capture time-series dependencies.

[0098] Furthermore, using the different data sources (voltage, current, power, environmental data, and user behavior data) processed in the previous steps, these expanded time-series features are now integrated into a first high-dimensional feature matrix F(t). The specific construction method is as follows:

[0099] F(t) = [Z V (t),Z I (t),Z P (t),Z T (t),Z H (t),Z cosφ (t),Z L (t)]

[0100] Among them, Z V (t) represents the time-series characteristics of the voltage, Z I (t) represents the current, Z P (t) represents power, Z T (t) and Z H (t) represent the temporal characteristics of temperature and humidity, respectively, Z C0sφ (t) represents the time-series characteristics of the load power factor, Z L (t) represents the timing characteristics of the electricity consumption pattern.

[0101] The final output is the first high-dimensional feature matrix F(t), which fuses multi-source data together in time series to form an input matrix containing rich time-series and multi-dimensional information, which is then used by the subsequent deep learning model. The construction of this feature matrix is ​​specifically designed for the complexity of medium-voltage distribution networks, ensuring that it not only incorporates electrical parameters but also the multi-dimensional influences of the environment and user behavior, greatly enhancing the accuracy of load forecasting.

[0102] This step, through five sub-steps (data collection, synchronization and normalization, noise filtering, feature expansion, and data fusion), provides a structured and rich input data matrix for subsequent model training and load prediction. In particular, through adaptive noise filtering and sliding window time-series feature generation, it addresses the problem that traditional methods cannot accurately capture the complex temporal behavior of power grid loads. The final output high-dimensional feature matrix F(t) serves as the input for subsequent steps, integrating multi-dimensional dynamic information of the power grid and effectively improving the model's training quality and prediction performance.

[0103] S2. Design an adaptive hierarchical clustering algorithm to dynamically identify load types based on a high-dimensional feature matrix. Then, extract the time-series features of the load and the autocorrelation features from the time-series features from the load types, and construct a second high-dimensional feature matrix from the time-series features and the autocorrelation features.

[0104] The input is the high-dimensional feature matrix F(t) generated in step 1, which has the following form:

[0105] F(t) = [Z V (t),Z I (t),Z P (t),Z T (t),Z H (t),Z cosφ (t),Z L (t)]

[0106] These data include electrical characteristics (such as voltage Z) V (t), Current Z I (t), power Z P (t)), and environmental parameters (such as temperature Z)T (t), humidity Z H (t) and user behavior data (such as load power factor Z) cosφ (t), User electricity consumption mode Z L (t)). These features, having undergone preprocessing in step 1, accurately reflect the load changes of the power grid at different points in time.

[0107] Specifically, in medium-voltage distribution networks, different types of loads (such as industrial loads, commercial loads, and residential loads) exhibit significantly different characteristics, and these characteristics change over time. Therefore, traditional static classification methods cannot accurately identify these load types. This invention designs an adaptive hierarchical clustering algorithm that can dynamically identify load types by combining electrical characteristics, environmental characteristics, and electricity consumption behavior characteristics.

[0108] Furthermore, due to the time-varying characteristics of the load, the present invention segments the feature matrix F(t) into segments with fixed windows to form a segmented matrix S(t) containing the past n time steps, in order to capture the short-term fluctuations and long-term trends of the load.

[0109] S(t)=[F(t),F(t-1),…,F(tn)]

[0110] This time-series segmentation processing method can combine the historical behavior of the load, making it particularly suitable for the fluctuating characteristics of loads in medium-voltage distribution networks, such as peak industrial loads and periodic changes in commercial loads.

[0111] Furthermore, the feature matrix S(t) is a high-dimensional matrix, and direct clustering would lead to interference from redundant information. Therefore, this invention uses principal component analysis (PCA) to reduce the dimensionality of S(t), retaining 99% of the variance information. After dimensionality reduction, the feature vector R(t) is generated:

[0112] R(t) = PCA(S(t))

[0113] By reducing dimensionality, this invention effectively reduces the computational complexity of high-dimensional matrices while retaining the most critical information about the load. This provides a concise and essential feature representation for processing multidimensional data (such as electrical and environmental data) in load characteristics.

[0114] Furthermore, based on the dimensionality-reduced feature vector R(t), this invention introduces an adaptive hierarchical clustering algorithm. This algorithm ensures accurate classification of load types by adjusting the clustering hierarchy. The clustering process performs initial clustering of R(t) using a distance metric (such as Euclidean distance) and adjusts the clustering hierarchy according to temporal characteristics, ultimately outputting the load classification label C(t):

[0115] C(t) = AdaptiveCluster(R(t))

[0116] Here, C(t) represents the load type (e.g., industrial, commercial, residential) at time t. Adaptive hierarchical clustering can dynamically adapt to fluctuations in load type by continuously optimizing the clustering results, and performs particularly well during periods of drastic load changes.

[0117] Furthermore, after identifying the load type, this invention extracts the time-series features of the load so that subsequent load prediction models can understand the short-term and long-term changes of different load types. This invention uses a sliding window and an autocorrelation function to extract the time-series features of the load type.

[0118] Furthermore, for each load type C(t), the features are expanded using a sliding window method to form a short-term time series feature vector T. C (t):

[0119] T C (t)=[F C (t),F C (t-1),…,F C (tn)]

[0120] By using a sliding window expansion, the characteristics of short-term fluctuations under load type C(t) are captured, especially for high-frequency fluctuating loads such as industrial loads. This ensures that subsequent forecasting models can accurately predict short-term load changes (such as instantaneous peak loads in industrial equipment).

[0121] Furthermore, in order to better capture the periodic changes in load, this invention calculates the autocorrelation function A under load type C(t). C (t), used to identify the periodic characteristics of long-term loads:

[0122]

[0123] Autocorrelation feature A C (t) can help this invention capture the periodic behavior of loads, especially the low-frequency fluctuations of residential loads and the high-frequency fluctuations of industrial loads. This feature extraction method enhances the performance of subsequent load prediction models under periodic load conditions.

[0124] Furthermore, after load type identification and time series feature extraction, the output second feature matrix E(t) contains time series features and autocorrelation features of different load types, providing rich input for subsequent load prediction models.

[0125] E(t) = [T] C (t),A C (t)]

[0126] Here, T C(t) is the sliding window timing characteristic for each load type, A C (t) represents the autocorrelation characteristic. These data help the model understand the short-term fluctuations and long-term cyclical changes of different loads, thereby enhancing the accuracy and robustness of predictions.

[0127] S3. Construct a time-series multi-level recurrent neural network, using the second high-dimensional feature matrix as input, and perform load prediction by combining short-term and long-term features through recursive structure and regularization methods.

[0128] Specifically, load variations in medium-voltage distribution networks are extremely complex, with different types of loads exhibiting significant differences across different time scales. Existing load forecasting methods often struggle to simultaneously address the short-term fluctuations of industrial loads and the periodic characteristics of residential loads. Therefore, this invention designs a Temporal Multilevel Recurrent Neural Network (T-MRNN) to address this complexity. Through an innovative recurrent structure and regularization method, it combines short-term and long-term characteristics for load forecasting.

[0129] First recurrent network: processes the short-term temporal features T of the input. C (t). This layer of the network is specifically designed to capture short-term load fluctuations, making it particularly suitable for handling frequent fluctuations in industrial loads and other similar conditions. The model formula is:

[0130] h1(t)=RNN1(T C (t))+λ1·reg(T C (t))

[0131] Where h1(t) is the hidden state of the first recurrent network, λ1 is the regularization coefficient, and reg(T) is the hidden state of the first recurrent network. C (t) is a specially designed regularization term. The regularization term is designed to address common sudden load changes in the power grid, such as load spikes caused by the start-up and shutdown of industrial equipment. It incorporates a local smoothing regularization term to avoid the model overfitting to abnormal fluctuations.

[0132] The formula for the local smoothing regularization term is as follows:

[0133]

[0134] This regularization term restricts sharp changes in short-term load characteristics, preventing the model from overfitting local fluctuations when industrial loads fluctuate frequently, and ensuring that the model can capture the overall trend of load changes.

[0135] The second recurrent network combines the hidden state h1(t) and the autocorrelation feature A from the first layer. C (t) Handling long-term load trends. This layer is specifically designed to handle periodic features, ensuring the model captures long-term load dependencies. The model formula is:

[0136] h2(t)=RNN2(h1(t),A C (t))+λ2·reg(A C (t))

[0137] Where h2(t) is the hidden state of the second recursive network, λ2 is the regularization coefficient, and reg(A C (t) is the long-term periodicity regularization term, used to suppress overfitting to periodic oscillations and ensure that load periodic fluctuations are correctly captured rather than amplified. The periodicity regularization term is designed as follows:

[0138]

[0139] This regularization term suppresses the model's fit to transitional periodic changes by comparing the autocorrelation characteristics at the current time step t with those at the previous m time steps. This regularization helps the model better capture long-term fluctuations in cyclical loads, such as the cyclical electricity consumption behavior in residential loads.

[0140] This hierarchical recursive network design, by combining short-term fluctuation handling with long-term trend capture, can simultaneously adapt to both short-term sudden changes and long-term stable changes in the load of the power grid. In particular, by introducing a regularization term to suppress the fitting of abnormal load fluctuations and transition period changes, the model can provide more stable load forecasts.

[0141] Furthermore, after processing through two layers of recursive networks, the final hidden state h2(t) is passed to a fully connected layer for mapping, generating the load prediction value P(t), as shown in the following formula:

[0142] P(t) = W p ·h2(t)+b p

[0143] Among them, W p It is the weight matrix of the fully connected layer, b p This is the bias term, where P(t) represents the predicted load value at future time step t. Through this prediction structure, the model can effectively convert multi-timescale features into specific load values. In particular, the model can generate stable and accurate prediction results for both short-term fluctuations in industrial load and long-term periodicity in residential load.

[0144] Meanwhile, this invention adds a load allocation constraint term to the prediction output layer to ensure that the predicted load meets the grid capacity limit. This constraint term C(t) is added to the output layer in matrix form:

[0145]

[0146] Here, P i (t) is the predicted value of the i-th type of load, Capacityi This refers to the corresponding distribution network capacity. This constraint is used to control the total predicted load to not exceed the maximum capacity ratio of the distribution network, thereby ensuring that the predicted load can be reasonably distributed under the constraints of network operation.

[0147] Furthermore, during training, this invention uses historical load data for supervised learning. To optimize the model, this invention employs a weighted mean squared error (WMSE) loss function, assigning different weights to errors of different load types to ensure accurate optimization for both high-frequency fluctuations and periodic loads. The formula is as follows:

[0148]

[0149] Where Y(t) is the actual load, P(t) is the predicted load, and α is a weighting factor dynamically adjusted according to the load type. Through a weighted loss function, the model can better adapt to error adjustments for different load types, especially for the different needs of industrial and residential loads.

[0150] Furthermore, the trained T-MRNN model ultimately generates a load prediction value P(t). This result not only reflects short-term fluctuations in the power grid but also incorporates long-term periodic changes, providing accurate data support for subsequent energy management and load scheduling. The prediction result has also been adjusted for load allocation constraints to ensure it conforms to the actual capacity limitations of the power grid and is practically operable.

[0151] S4. Calculate the current capacity ratio of the power grid based on the load forecast, and perform preliminary optimization on the current capacity ratio to obtain a preliminary optimized load forecast.

[0152] Specifically, the capacity-to-load ratio (R(t)) of a medium-voltage distribution network is the ratio between the load at each node in the network and its available capacity, used to measure whether the network is under normal load conditions. To calculate the capacity-to-load ratio, this invention combines the predicted value P for different load types. i (t) and the available capacity of each node. i A formula specifically designed for real-time capacity ratio calculation is proposed. The formula for calculating the capacity ratio is:

[0153]

[0154] Among them, P i (t) is the predicted value of the i-th type of load, Capacity iR(t) represents the capacity of the grid node corresponding to this load type. R(t) is the total capacity-to-load ratio at time t. This calculation ensures that the grid load does not exceed the node capacity and can reflect the grid's operating status in real time. If R(t) ≥ 1, it indicates that the grid load is overloaded and load scheduling is required through the energy management system; while if R(t) < 1, it indicates that the grid load is at a normal level.

[0155] Furthermore, due to the significant fluctuations and periodicity of loads in medium-voltage distribution networks, energy allocation directly based on load forecasting results may lead to local node overload. To address this issue, this invention designs an adaptive load balancing optimization model that adjusts the load allocation of each node to ensure the overall stability of the power grid.

[0156] First, based on the capacity ratio R of each node i (t), load transfer is performed on nodes whose load exceeds a set threshold. To this end, this invention introduces a node load adjustment term ΔP. i (t), which is dynamically adjusted based on the current capacity ratio of each node and the load of its neighboring nodes.

[0157] The node load adjustment formula is:

[0158]

[0159] Where α is the load balancing adjustment coefficient, used to control the load distribution speed. β is the neighbor node adjustment coefficient, used to adjust the load difference with neighboring nodes. N(i) is the set of nodes adjacent to node i, and |N(i)| is the number of neighboring nodes. P j (t) represents the predicted load value of the adjacent node j. This load adjustment formula allows for the reasonable distribution of overloaded node loads to adjacent nodes with lighter loads, ensuring a balanced distribution of the overall power grid load. Especially when industrial loads are excessively high, this scheme can dynamically adjust the load, preventing localized overloads that could lead to power outages.

[0160] Furthermore, through the aforementioned capacity-to-load ratio calculation and energy distribution adjustment, the real-time optimization scheme can ensure the stability of the power grid under various complex load conditions. The entire optimization process is not only based on the load and capacity calculation of each node, but also makes further dynamic adjustments based on the load information of adjacent nodes to ensure rapid balancing when local loads are overloaded.

[0161] To more accurately reflect the actual operation of the power grid, this invention introduces a power grid dynamic adjustment function D(t) to describe the dynamic rate of change of the current capacity-to-load ratio in the power grid:

[0162]

[0163] This formula represents the impact of load change on the capacity ratio at each time step. By observing D(t), this invention can identify areas in the power grid where the load changes rapidly and perform load scheduling in advance based on this trend. If the value of D(t) increases positively, it indicates that the load is growing rapidly and further optimization is needed; if the value decreases, it indicates that the power grid is in a load-balanced state.

[0164] Furthermore, in practical energy management systems, this invention performs real-time optimization based on capacity-to-load ratio calculation and node adjustment to ensure the stability of the entire power grid. To improve the system's adaptability, this invention designs a feedback mechanism that feeds back the optimized power grid state to the next prediction model. The specific steps are as follows:

[0165] The capacity ratio R(t) of each node is fed back to the load prediction model to adjust the prediction weights of future loads.

[0166] The overloaded nodes are gradually adjusted to prevent local overload from recurring in the next time period.

[0167] Through this feedback mechanism, the entire optimization process forms a closed loop, enabling the calculation of the load ratio to be linked with the energy management system in real time, ensuring the system's sensitivity and adaptability to load fluctuations.

[0168] S5. Based on the preliminary optimized load prediction, a deep learning model is constructed to further optimize the energy management strategy, and the model parameters of the deep learning model are adjusted to adapt to the actual operating conditions.

[0169] Specifically, to achieve load balancing among different nodes in the power grid, this invention designs an adaptive load optimization mechanism. This mechanism combines the load prediction and adjustment information from step 4 to correct the model's prediction results in real time, thereby improving overall scheduling efficiency. The optimization objective is to minimize load imbalance in the power grid while ensuring that the load of each node is within its capacity range. To this end, this invention introduces an adaptive optimization function based on the capacity-to-load ratio, as shown in the following formula:

[0170]

[0171] Among them, R i (t) is the capacity ratio of the i-th node. P i (t) is the predicted load value for this node. Capacity i This is the capacity of the node. ΔP iO(t) represents the node load adjustment. γ is the weighting coefficient of the load balancing adjustment term, used to control the relative importance of capacity ratio optimization and node load adjustment. The adaptive optimization mechanism ensures load balance by minimizing O(t), while dynamically adjusting the load forecast value to make the actual operating state of the power grid closer to the load forecast result, reducing load imbalance between nodes. This optimization process can handle load changes between nodes in real time, especially under conditions of drastic industrial load fluctuations, and has good adaptability to load scheduling.

[0172] Furthermore, based on the above optimization results, this invention adaptively updates the deep learning model to improve its accuracy and stability in predicting future loads. To capture nonlinear load changes, especially to address dynamic adjustments to unbalanced loads, this invention introduces an adaptive regression update mechanism for load changes:

[0173] First, calculate the current load prediction error ∈(t), as follows:

[0174]

[0175] Among them, Y i (t) represents the actual load value. By calculating the prediction error of each node, this invention can adjust the weights of the model.

[0176] To improve the model's adaptability to drastic load fluctuations, this invention designs an update mechanism with weight decay, as shown in the following formula:

[0177]

[0178] Where ΔW(t) is the adjustment amount of the model weights. η is the learning rate, used to control the speed of weight updates. It is the load ratio term, used to appropriately reduce weight adjustments when the load is close to the node capacity.

[0179] Through this adaptive update mechanism, the model can dynamically adjust its weights based on real-time load prediction errors. Especially when the load is too high, the model can gradually reduce its sensitivity to the current load, avoiding overfitting. Simultaneously, when the load is low, the model can more sensitively capture load changes, ensuring prediction accuracy.

[0180] Furthermore, to enhance the adaptability of power grid load dispatching, this invention designs a feedback closed-loop control mechanism that feeds back the model's predicted output and optimization results to the load dispatching system. The specific steps are as follows:

[0181] Load forecasting feedback: The load forecasting error ∈(t) and the capacity ratio R(t) are fed back to the scheduling system, and the load forecasting model for the next round is dynamically adjusted to better reflect the actual load changes.

[0182] Load balancing optimization feedback: The adjustment result ΔP from each load balancing iteration. i (t) is fed back to the deep learning model, dynamically adjusting the model parameters so that the model can more accurately reflect the actual situation of load scheduling.

[0183] Through this feedback loop, the power grid system can self-adjust according to the actual operating status in each cycle, improve the accuracy of load forecasting and the efficiency of load distribution, and ensure that the power grid remains stable under various load fluctuations.

[0184] This step, through a combination of adaptive optimization and model updates, enables real-time optimization and scheduling under complex load scenarios, demonstrating high adaptability, particularly in industrial and commercial load scenarios with drastic load changes. By optimizing the capacity ratio of each node in the power grid and combining it with adaptive updates of the deep learning model, the solution achieves dynamic load allocation and high-precision load prediction, further improving the operating efficiency of the power grid.

[0185] S6. Monitor the operation of the medium-voltage distribution network in real time, and use reinforcement learning models to optimize deep learning models and energy management strategies.

[0186] Specifically, addressing the uncertainty and volatility of power grid load in the patent scenario, this invention designs a deep reinforcement learning mechanism specifically adapted to load fluctuations in medium-voltage distribution networks. This mechanism can intelligently adjust the load distribution of the power grid, ensuring continuous optimization of power grid operation even during drastic load changes. The design of this invention not only considers load prediction errors but also optimizes power grid scheduling decisions through real-time adjustments to load distribution.

[0187] At each time step, the system state S(t) includes the current capacity ratio R of each node in the power grid. i (t), prediction error ∈ (t), and load adjustment ΔP between adjacent nodes. i (t). These state variables comprehensively reflect the real-time operating status of the power grid and serve as the basis for reinforcement learning to make optimal scheduling decisions. The state is defined as:

[0188] S(t) = {R1(t), R2(t), ..., R} n (t),∈(t),ΔP1(t),…,ΔP n (t)}

[0189] Unlike traditional power grid dispatching, the load changes in the scenario described in this invention have greater randomness. Therefore, state S(t) is specifically designed to adjust the load of adjacent nodes by ΔP.i The sensitivity of (t) is adjusted so that the model can monitor changes in local node loads in real time and ensure overall grid load balance.

[0190] For each node's state, the reinforcement learning model can choose actions including adjusting load allocation and updating model parameters. Unlike common power grid dispatching, this invention introduces a load allocation mechanism based on regional optimization, where action A(t) considers not only the load of a single node but also its interactions with neighboring nodes, ensuring regional load balance. The action vector is defined as:

[0191] A(t)={ΔP i (t),ΔW(t)}

[0192] By dynamically adjusting the node load distribution ΔP i With model weight update ΔW(t) and model weight update ΔW(t), the reinforcement learning model can flexibly adapt to changes in load in different regions and intelligently optimize load scheduling.

[0193] To address the load volatility in the patent scenario, this invention designs a reward function closely integrated with grid energy efficiency, load balancing, and prediction accuracy. This reward function adjusts the capacity ratio R... i The deviation of (t), the prediction error ∈ (t), and the load adjustment ΔP i (t) A comprehensive evaluation is conducted to ensure that the power grid maintains load balance while operating at high efficiency. The reward function is defined as:

[0194]

[0195] Among them, R i (t) represents the load-capacity ratio of node i, with a target value close to 1, indicating that the node's load is within its capacity. α is a coefficient used to adjust the smoothness of load adjustment, preventing over-adjustment that could lead to instability. β is used to control the impact of prediction errors on the reward, ensuring continuous improvement in model prediction accuracy. This reward function guarantees load balance and efficiency optimization during grid dispatching, particularly addressing the frequent load fluctuations in medium-voltage distribution networks. By penalizing excessive load adjustments and prediction errors, it enhances the model's adaptive dispatching capabilities.

[0196] Furthermore, this invention employs a Deep Q-Learning algorithm (DQN) for optimization. Through a reinforcement learning mechanism, the system can continuously optimize the load scheduling strategy during real-time operation. The specific process is as follows:

[0197] The Q-value is calculated using state S(t) and action A(t), reflecting the expected cumulative reward of taking a certain action in a specific state. The Q-value update formula is as follows:

[0198]

[0199] Here, η is the learning rate, used to control the magnitude of Q-value updates. γ is the discount factor, balancing the impact of current rewards against future gains. R(t) is the current reward signal, representing the system's performance under the current load schedule. This Q-value update process allows the system to continuously optimize its scheduling strategy, ensuring the optimal resource allocation scheme is found amidst complex load fluctuations.

[0200] By employing trial and error and leveraging balance through reinforcement learning, the model selects the action A(t) that maximizes the Q value at each time step, and then performs load adjustment ΔP. i The model weights are updated using ΔW(t) and model weights. This adaptive decision-making approach ensures that the system can adjust its strategy according to real-time load fluctuations, enabling the power grid to operate efficiently even under complex conditions.

[0201] Furthermore, to address sudden load changes and long-term fluctuations in the power grid, this invention designs a continuous optimization feedback mechanism based on reinforcement learning. At the end of each decision cycle, the system will input the load prediction error ∈(t) and the load adjustment amount ΔP. i (t) is fed back to the reinforcement learning model, enabling it to dynamically adjust its scheduling decisions for the next round based on the latest state. The specific feedback process is as follows:

[0202] The results of each round of load scheduling are fed back to the reinforcement learning model to adjust the state evaluation and reward function for the next round, ensuring that the system can continuously improve the scheduling strategy based on historical data.

[0203] After each round of decision-making, the reinforcement learning model adjusts the node load based on feedback, allowing the load to tend towards equilibrium within future time steps. The system can adaptively adjust based on real-time changes in power grid data, avoiding excessive fluctuations and the impact of sudden load surges on the power grid.

[0204] For the medium-voltage distribution network in the patent scenario, load fluctuations and regional load imbalances are key issues that this invention aims to address. By combining reinforcement learning mechanisms with adaptive load adjustment, this invention enables high-frequency load scheduling and optimization, allowing the system to cope with complex load environments with multiple nodes and ensuring stable operation of the power grid even under severe load fluctuations. This invention specifically designs a regional load adjustment mechanism to address the load imbalance problem, ensuring dynamic adjustment of load distribution between adjacent nodes and enhancing system flexibility.

[0205] These are merely some preferred embodiments of the present invention, and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the processes of the above embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network, characterized in that, The method includes the following steps: S1. Collect multi-source data from the medium-voltage distribution network and construct the first high-dimensional feature matrix based on the multi-source data; S2. Design an adaptive hierarchical clustering algorithm to dynamically identify load types based on a high-dimensional feature matrix. Then, extract the time-series features of the load and the autocorrelation features from the time-series features from the load types, and construct a second high-dimensional feature matrix from the time-series features and the autocorrelation features. S3. Construct a time-series multi-level recurrent neural network, using the second high-dimensional feature matrix as input, and perform load prediction by combining short-term and long-term features through recursive structure and regularization methods. S4. Calculate the current capacity ratio of the power grid based on the load forecast, and perform preliminary optimization on the current capacity ratio to obtain a preliminary optimized load forecast. S5. Based on the preliminary optimized load prediction, a deep learning model is constructed to further optimize the energy management strategy, and the model parameters of the deep learning model are adjusted to adapt to the actual operating conditions. S6. Monitor the operation of the medium-voltage distribution network in real time, and use reinforcement learning models to optimize deep learning models and energy management strategies.

2. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 1, characterized in that, The multi-source data includes electrical data, environmental data, and user behavior data; The electrical data includes voltage, current, and power; The environmental data refers to environmental changes. The user behavior data refers to load fluctuations caused by different users' electricity consumption patterns. These load fluctuations include load power factor and electricity consumption characteristics. The construction of the first high-dimensional feature matrix based on multi-source data specifically includes: The multi-source data is preprocessed, including data synchronization and normalization, noise filtering and outlier handling, and feature expansion and time-series feature generation. Finally, time-series features are generated and integrated into a first high-dimensional feature matrix.

3. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 1, characterized in that, The adaptive hierarchical clustering algorithm is designed to dynamically identify load types based on a high-dimensional feature matrix, specifically including: The first high-dimensional feature matrix will be segmented according to a fixed window to form a matrix containing past features. Piecewise matrix at each time step Then, principal component analysis was used to analyze the piecewise matrix. Dimensionality reduction to generate feature vectors Finally, an adaptive hierarchical clustering algorithm is introduced to perform feature vector analysis using a distance metric. Perform initial clustering and adjust the clustering hierarchy according to time series characteristics to finally output the load type. ; The step of extracting the time-series features of the load and the autocorrelation features from the time-series features from the load type, and constructing a second high-dimensional feature matrix from the time-series features and the autocorrelation features, specifically includes: For each load type The sliding window method is used to expand its features to form time-series features for each load type. Simultaneously calculate load type Autocorrelation function under Identify the periodic characteristics of long-term loads, and ultimately determine the autocorrelation function. and timing characteristics of each load type Reconstruct the second high-dimensional feature matrix .

4. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 1, characterized in that, The temporal multi-level recurrent neural network includes a first recurrent network and a second recurrent network; The first layer of the recursive network processes the temporal characteristics of each load type, as shown below: ； in, This is the hidden state of the first layer of the recursive network. These are the regularization coefficients of the local smoothing regularization term. It is a specially designed local smoothing regularization term. It is the first layer of recursive network. It is a temporal feature; wherein, the local smoothing regularization term It is expressed as follows: ； Where n is the total number of windows, and i is the window size. It is time t; The second recurrent network combines the hidden states and autocorrelation features of the first layer to process the long-term trend of the load, as shown below: ； in, It is the hidden state of the second-layer recursive network. It is an autocorrelation feature. It is the regularization coefficient of the long-term periodic regularization term; It is a long-term periodic regularization term, represented as follows: ； Where m is the total number of time steps, and k is the k time steps; The final hidden state is obtained through the output of the aforementioned temporal multi-level recurrent neural network. It is passed to the fully connected layer for mapping to generate load prediction values. , means as follows: ； in, It is the weight matrix of the fully connected layer. It is a bias term. This represents the predicted load value at time t; In S3, the method further includes adding a load allocation constraint term to the prediction output layer of the time-series multi-level recurrent neural network to constrain the load to meet the grid capacity, as shown below: ； in, It is a constraint term. It is the predicted load value for node i. It is the corresponding distribution network capacity.

5. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 4, characterized in that, During the training process of the aforementioned temporal multi-level recurrent neural network, supervised learning is performed using historical load data, and a weighted mean square error loss function is employed to assign different weights to errors of different load types, as shown below: ； in, This is the actual load. It is the load forecast value at time t. It is a weighting factor that is dynamically adjusted based on the load type.

6. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 1, characterized in that, The preliminary optimization of the current capacity ratio to obtain a preliminary optimized load prediction specifically includes: Construct an adaptive load balancing optimization model based on the capacity ratio of each node. Load shifting is performed on nodes whose load exceeds a set threshold, as shown below: ； in, This is a node load adjustment item, which is dynamically adjusted based on the current capacity ratio of each node and the load status of its neighboring nodes; It is the load balancing adjustment factor, used to control the load distribution speed; It is the adjacent node adjustment coefficient, used to adjust the load difference with adjacent nodes; It is with nodes The set of adjacent nodes, It is the number of adjacent nodes; Adjacent nodes The predicted load value; It is a node The capacity of the power grid node corresponding to the load type; It is a node The predicted load value; t is the time t; Simultaneously, a dynamic adjustment function for the power grid is introduced. This is used to describe the dynamic rate of change of the current power grid capacity ratio, and is expressed as follows: ； According to the current power grid dynamic adjustment function The value is used to identify areas of load change in the power grid and to schedule loads in advance based on the degree of change. Among them, the capacity ratio of each node It is the ratio between the load on each node in the power grid and its available capacity, used to measure whether the power grid is under normal load conditions, and is expressed as follows: ； in, It is the predicted load of node i; It is a node The capacity of the power grid node corresponding to the load type; It is the total capacity ratio at time t.

7. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 1, characterized in that, S5 specifically includes: S5.1 An adaptive load optimization mechanism is designed to perform real-time correction of the capacity ratio of each node, ensuring that the capacity ratio of each node is within its capacity range; wherein, the adaptive load optimization mechanism minimizes an adaptive optimization function based on the capacity ratio. To ensure load balance, the load forecast is dynamically adjusted so that the actual operating state of the power grid is as close as possible to the load forecast result; the adaptive optimization function It is expressed as follows: ； in, It is the first The load capacity ratio of the node; This is the predicted load value for that node; This is the capacity of the node; It is the node load adjustment amount; It is the weighting coefficient of the load balancing adjustment item, used to control the relative importance of capacity ratio optimization and node load adjustment; S5.2 An adaptive regression update mechanism is designed to adapt the deep learning model to changes in load. S5.

3. Implement closed-loop feedback for the deep learning model, feeding back the load prediction error and capacity ratio to the scheduling system, and dynamically adjusting the load prediction model for the next round. At the same time, feed back the node load adjustment amount for each load allocation to the deep learning model, and dynamically adjust the model parameters of the deep learning model.

8. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 7, characterized in that, S5.2 specifically includes: First, calculate the current load prediction error. , means as follows: ； in, This is the actual load value; The design of an update mechanism with weight decay allows deep learning models to have the best adaptability when the load fluctuates drastically, as shown below: ； in, It is the adjustment amount of the model weights. It is the learning rate, used to control the speed at which weights are updated. It is the load ratio term, used to mitigate weight adjustments when the load approaches the node's capacity.

9. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 8, characterized in that, The state of the reinforcement learning model includes the capacity ratio of each node in the current power grid, the prediction error, and the load adjustment between adjacent nodes; the actions include dynamically adjusting the node load distribution and updating the model weights of the deep learning model; the reward function... It is expressed as follows: ； in, It is a node The load ratio is targeted to be close to 1, indicating that the load on a node is within its capacity. It is a coefficient used to regulate the smoothness of load adjustment, preventing instability caused by over-adjustment; It is used to control the impact of prediction errors on rewards and ensure that the model's prediction accuracy continues to improve.

10. The method for calculating the capacity-to-load ratio of a medium-voltage distribution network based on a deep neural network according to claim 9, characterized in that, The reinforcement learning model is optimized using a deep Q-learning algorithm, through state... and actions Calculate the Q-value at the current moment, which reflects the expected cumulative reward of taking a certain action in a specific state; the Q-value update formula is as follows: ； in, It is the learning rate, used to control the magnitude of Q-value updates. It is a discount factor that weighs the impact of current rewards against future benefits. It is the current reward function, representing the operating effect under the current load scheduling.

Images