A grid control method and system based on deep learning and photovoltaic power prediction

Abstract

A power grid control method and system based on deep learning and photovoltaic power prediction, the method comprising: acquiring photovoltaic data; constructing a photovoltaic power prediction network; training the photovoltaic power prediction network and deploying it on the device side; predicting actual photovoltaic power to obtain a photovoltaic power generation power prediction value; an automatic power generation control system calculating the output of other power generation equipment that needs to be adjusted when the photovoltaic power generation power prediction value changes based on the photovoltaic power generation power prediction value, the current power grid load data, and the output data of other power generation equipment, and adjusting the output to maintain the active power balance of the power grid; an automatic voltage control system adjusting the reactive power based on the photovoltaic power generation power prediction value and the voltage data monitored in real time by the power grid, so that the power grid voltage is stabilized within an allowable range. The present invention realizes closed-loop management from power generation power prediction to real-time power grid control, effectively improving the power grid's ability to cope with fluctuations in photovoltaic power generation output and ensuring the safe and stable operation of the power grid.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid regulation, and in particular to a power grid regulation method and system based on deep learning and photovoltaic power prediction. Background Art

[0002] As the global energy transition continues, photovoltaic power generation, driven by its numerous advantages, such as clean and renewable energy, is rapidly expanding its use in power supply systems and being integrated into the grid on a large scale. However, the inherent characteristics of photovoltaic power generation present a series of difficult challenges, posing a serious threat to the stable operation of the power supply system. Traditional power supply systems have long employed a scheduling model that prioritizes thermal power generation with photovoltaic power generation as a supplement. As the proportion of photovoltaic power generation in the power supply system continues to increase, this traditional scheduling model has gradually exposed its limitations. Thermal power generation units have relatively slow regulation speeds, making them unable to quickly respond to rapid fluctuations in photovoltaic power generation. Furthermore, traditional scheduling models are difficult to adapt to the complex power supply scenarios associated with the integration of a high proportion of renewable energy sources. Summary of the Invention

[0003] The present invention provides a grid control method and system based on deep learning and photovoltaic power prediction to solve the technical problems mentioned in the background technology.

[0004] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0005] The present invention provides a grid control method based on deep learning and photovoltaic power prediction, comprising the following steps:

[0006] S1. Acquire and preprocess photovoltaic data; construct a photovoltaic power prediction network CH-BiLSTM, which includes a sequentially connected one-dimensional convolutional neural network, a pooling layer, a hierarchical attention mechanism, a bidirectional long short-term memory network, and a fully connected layer;

[0007] S2. Train the photovoltaic power prediction network CH-BiLSTM to obtain the trained photovoltaic power prediction network CH-BiLSTM and deploy it on the device side; use the photovoltaic power prediction network CH-BiLSTM on the device side to perform actual photovoltaic power prediction and obtain the photovoltaic power prediction value. ;

[0008] S3, automatic power generation control system AGC based on the photovoltaic power generation power prediction value , current grid load data and output data of other power generation equipment Calculate the predicted value of photovoltaic power generation In case of changes, other power generation equipment needs to adjust their output and make adjustments to maintain the active power balance of the grid;

[0009] S4, automatic voltage control system AVC according to the photovoltaic power generation power forecast value And the voltage data of the power grid in real time monitoring Regulating reactive power , in order to change the reactive power distribution in the power grid and stabilize the grid voltage within the allowable range.

[0010] On the other hand, the present invention also provides a power grid control system based on deep learning and photovoltaic power prediction, including a power grid, which is controlled by the above-mentioned power grid control method.

[0011] Beneficial effects of the present invention:

[0012] 1. The present invention discloses a photovoltaic power prediction network CH-BiLSTM. The photovoltaic power prediction network CH-BiLSTM is designed with a hierarchical attention mechanism. The hierarchical attention mechanism can more clearly identify and retain key moments through discrete selection operations. At the same time, the hierarchical attention mechanism only retains the most representative The information weight distribution is clearer and the interference of redundant information and noise interference is effectively reduced, thereby improving the robustness of the photovoltaic power prediction network CH-BiLSTM;

[0013] In addition, after extracting local features, the hierarchical attention mechanism in the present invention re-screens key local information and combines it with global timing information, enabling the photovoltaic power prediction network CH-BiLSTM to more comprehensively capture the dynamic changes of data.

[0014] 2. The present invention collaboratively controls the power grid through the automatic generation control system AGC and the automatic voltage control system AVC. This collaborative control mechanism breaks the situation where traditional prediction models and power supply system scheduling are independent of each other, and realizes closed-loop management from power generation power prediction to real-time grid regulation, effectively improving the power supply system's ability to cope with fluctuations in photovoltaic power generation output and ensuring the safe and stable operation of the power supply system.

[0015] 3. This invention also discloses a photovoltaic power generation prediction system based on CH-BiLSTM. Compared with traditional photovoltaic power generation prediction systems, this invention responds more quickly to abnormalities and sudden changes. It also offers higher prediction accuracy than traditional statistical and physical models. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of the power grid control method in the present invention;

[0017] Figure 2 This is a structural block diagram of the photovoltaic power prediction network CH-BiLSTM in the present invention;

[0018] Figure 3 This is the collaborative control mechanism of the power supply system in the present invention. DETAILED DESCRIPTION

[0019] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The accompanying drawings illustrate preferred embodiments of the present invention. However, the present invention may be implemented in many other forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present disclosure.

[0020] Reference Figures 1 to 3 , an embodiment of the present application provides a grid control method based on deep learning and photovoltaic power prediction, comprising the following steps:

[0021] S1. Acquire and preprocess photovoltaic data; construct a photovoltaic power prediction network CH-BiLSTM, which includes a sequentially connected one-dimensional convolutional neural network, a pooling layer, a hierarchical attention mechanism, a bidirectional long short-term memory network, and a fully connected layer;

[0022] S2. Train the photovoltaic power prediction network CH-BiLSTM to obtain the trained photovoltaic power prediction network CH-BiLSTM and deploy it on the device side; use the photovoltaic power prediction network CH-BiLSTM on the device side to perform actual photovoltaic power prediction and obtain the photovoltaic power prediction value. ;

[0023] Specifically, the photovoltaic power prediction network CH-BiLSTM uses an end-to-end training approach, with each module jointly optimized through backpropagation. Since photovoltaic data is often affected by insufficient data volume and noise interference, data augmentation strategies such as random cropping, rotation, translation, and noise injection are introduced to improve the generalization ability of the photovoltaic power prediction network CH-BiLSTM. Cross-validation and early stopping strategies are also used to prevent overfitting. Through these training strategies, the photovoltaic power prediction network CH-BiLSTM can maintain low prediction errors in the face of different weather conditions and load scenarios, and demonstrates higher accuracy and robustness than traditional models in short-term prediction tasks.

[0024] S3, automatic power generation control system AGC based on photovoltaic power generation power prediction value , current grid load data and output data of other power generation equipment Calculate the predicted value of photovoltaic power generation In case of changes, other power generation equipment needs to adjust their output and make adjustments to maintain the active power balance of the grid;

[0025] S4, automatic voltage control system AVC according to the photovoltaic power generation power forecast value And the voltage data of the power grid in real time monitoring Regulating reactive power , in order to change the reactive power distribution in the power grid and stabilize the grid voltage within the allowable range. The coordinated control mechanism of the power supply system (or power grid) in the present invention is shown in S3 and S4.

[0026] In some embodiments, the step of obtaining photovoltaic data and preprocessing the photovoltaic data in S1 specifically includes the following steps:

[0027] S11. Obtain photovoltaic data including at least solar irradiance, atmospheric pressure, air temperature, and relative humidity;

[0028] S12. Normalize the photovoltaic data to eliminate the dimensional differences between data of different scales, making it easier for the photovoltaic power prediction network CH-BiLSTM to converge during training. Normalization preprocessing (such as Min-Max normalization or Z-score standardization) is used to adjust the eigenvalues ​​of the photovoltaic data to the same range to obtain the adjusted photovoltaic data.

[0029] S13: Encode additional information including at least season and time period and fuse it with the adjusted PV data to obtain preprocessed PV data. Normalization preprocessing not only ensures data consistency but also provides robust initial features for subsequent modules.

[0030] In some embodiments, a dynamic feature fusion strategy is introduced into the one-dimensional convolutional neural network (1dCNN), that is, by adaptively adjusting the feature weights within each time window, the features output by the one-dimensional convolutional neural network can capture local changes and suppress noise and redundant information to a certain extent;

[0031] The BiLSTM network includes two layers of LSTM (Long Short-Term Memory) artificial neural networks. LSTM adds gating units to the RNN (Recurrent Neural Network), which can adaptively retain or forget information, thus having better performance when processing large amounts of data.

[0032] The two-layer LSTM (long short-term memory) artificial neural network consists of a forward LSTM and a backward LSTM. The bidirectional processing capabilities of the bidirectional LSTM network help the CH-BiLSTM photovoltaic power prediction network understand sequential data and identify complex temporal features in energy consumption data. The forward LSTM processes the weighted features obtained from the hierarchical attention mechanism from beginning to end, while the backward LSTM processes the weighted features obtained from the hierarchical attention mechanism from end to beginning, thereby capturing more contextual relationships from past and future inputs. The bidirectional LSTM network allows gradients to propagate in both directions when processing data in both directions, thereby improving the stability of the CH-BiLSTM photovoltaic power prediction network and reducing the possibility of gradient vanishing, gradient dropout, and gradient explosion.

[0033] In some embodiments, the step S2 specifically includes the following steps:

[0034] S21, inputting the preprocessed photovoltaic data into a one-dimensional convolutional neural network to obtain multiple local features, and then inputting the multiple local features into a pooling layer to obtain a feature sequence;

[0035] S22, the hierarchical attention mechanism uses discrete selection operations to retain the previous time steps and convert them into binary indicator variables respectively. The binary indicator variables of time steps are concatenated in time order to obtain weighted feature representation;

[0036] S23. Use a bidirectional long short-term memory network to perform temporal modeling on the weighted feature representation to obtain a forward hidden state and a backward hidden state. The forward hidden state and the backward hidden state are then concatenated to obtain a final hidden state. The final bidirectional hidden state is linearly mapped through a fully connected layer to ultimately generate a predicted value of photovoltaic power.

[0037] S24. Construct a loss function based on the predicted value and the true value of the photovoltaic power, iterate S21 to S23, minimize the loss function, and adjust the weight of the photovoltaic power prediction network CH-BiLSTM during the iteration process to obtain the trained photovoltaic power prediction network CH-BiLSTM.

[0038] S25. Deploy the trained photovoltaic power prediction network CH-BiLSTM to the device side, and use the photovoltaic power prediction network CH-BiLSTM on the device side to predict the actual photovoltaic power and obtain the photovoltaic power prediction value. .

[0039] In some embodiments, the step S21 specifically includes the following steps:

[0040] S211, input the pre-processed photovoltaic data into a one-dimensional convolutional neural network, and the one-dimensional convolutional neural network uses a dynamic feature fusion strategy to extract local features of the pre-processed photovoltaic data to obtain multiple local features. , where local features Including short-term meteorological mutations, such as irradiance drop, temperature fluctuation and other information; power response mode, such as ramp rate, hysteresis effect and other information, local characteristics The calculation formula is:

[0041] ;

[0042] in, is the weight parameter of the convolution kernel at the jth position, b is the bias, is the activation function, is the convolution kernel size; Represents the irradiance and power of the time series data at time t+j;

[0043] S212, multiple local features Input into the pooling layer to obtain the feature sequence H, which is: ;

[0044] in, Represent the local features of the 1st, 2nd, and Tth time steps respectively.

[0045] In some embodiments, the step S22 specifically includes the following steps:

[0046] S221. Input the feature sequence H into the hierarchical attention mechanism and calculate the score of each time step based on the feature sequence H. The calculation formula is as follows:

[0047] ;

[0048] in, represents the score of the i-th time step; 、 Two learnable parameters for score calculation, is a trainable weight vector; Represents the feature vector of the i-th time step; Used to capture the contribution of different features to the judgment of key time steps; Used to determine the filtering priority of the time step;

[0049] S222, use discrete selection operation to convert the score of each time step into a binary indicator variable ;

[0050] S223. Preset a critical time step number , for all time steps according to Sort by size, with the highest score binary indicator variable corresponding to each time step is assigned a value of 1, and the other time steps are assigned a value of 0 to retain the one with the highest score time steps; in the above, the number of critical time steps The calculation formula is as follows:

[0051] ;

[0052] Where n represents the total number of time steps;

[0053] S224, the highest retention score The binary indicator variable corresponding to each time step Cascade in chronological order to form the weighted feature representation Z, which is calculated as follows:

[0054] ;

[0055] Among them, Concat(·) represents the feature concatenation operation.

[0056] In some embodiments, S23 specifically includes the following steps:

[0057] S231. Input the weighted feature representation Z into the forward long short-term memory network (LSTM) within the bidirectional long short-term memory network. The forward long short-term memory network (LSTM) processes the weighted feature representation Z from start to finish and calculates a forward hidden state during the processing. The forward hidden state represents a dynamic memory unit for historical weighted features. The gating mechanism selectively retains temporal patterns related to photovoltaic power prediction, such as cloud cover period and solar radiation change trend. The forward hidden state is used to provide a context-aware feature representation for the final prediction. The forward hidden state is calculated as follows:

[0058] ;

[0059] in, represents the forward hidden state; is the weighted feature output from the attention mechanism layer at the tth time step, It is t -1 time step forward hidden state, and is the forward cell state at the t-1th time step. The forward cell state represents the core memory unit of the forward LSTM, responsible for long-term storage and transmission of time series information when processing data in chronological order. It is the core mechanism for LSTM to achieve long-term memory. In photovoltaic forecasting, it is responsible for storing cross-time-step regularities (such as daily cycles and sudden weather events) in time series data such as irradiance and power. Represents the forward long short-term memory network LSTM;

[0060] S232. The backward long short-term memory network LSTM in the bidirectional long short-term memory network processes the weighted feature representation Z from the end to the beginning, and calculates the backward hidden state during the processing; the calculation formula of the backward hidden state is:

[0061] ;

[0062] in, Represents the backward hidden state; is the backward hidden state at the t+1th time step, is the backward cell state at the t + 1th time step, where the backward cell state is the internal state vector storing long-term memory in the forward LSTM and is updated in chronological order; Represents the backward long short-term memory network LSTM;

[0063] S233, concatenate the forward hidden state and the backward hidden state to obtain the final bidirectional hidden state; the calculation formula is:

[0064] ;

[0065] in, represents the final bidirectional hidden state;

[0066] S234. Input the final bidirectional hidden state into the fully connected layer, and after linear mapping of the fully connected layer, finally generate the predicted value of photovoltaic power.

[0067] In some embodiments, S3 specifically includes the following steps:

[0068] S31, the photovoltaic power generation power prediction value , current grid load data and output data of other power generation equipment Transmitted together to the automatic power generation control system AGC;

[0069] S32, calculate the photovoltaic power generation power prediction value based on the power balance principle of the automatic power generation control system AGC In case of changes, the output of other power generation equipment needs to be adjusted , as follows:

[0070] ;

[0071] in, is the photovoltaic power generation power at the previous moment, is the load power at the previous moment;

[0072] S33, the automatic power generation control system AGC adjusts the output of other power generation equipment according to the calculation Send adjustment instructions, such as adjusting the valve opening of thermal power units or the guide vane opening of hydropower units, to adjust the power generation output, maintain the active power balance of the power grid, and ensure that the power grid frequency is stable within the specified range.

[0073] In some embodiments, the S4 specifically includes the following steps:

[0074] S41, automatic voltage control system AVC according to the photovoltaic power generation power forecast value And the voltage data of the power grid monitored in real time Regulating reactive power ;

[0075] S42, automatic voltage control system AVC based on reactive power Adjust the reactive power compensation equipment or adjust the excitation current of the generator to change the reactive power distribution in the power grid and stabilize the grid voltage within the allowable range.

[0076] In some embodiments, the reactive power in S41 The calculation formula is as follows:

[0077] ;

[0078] in, is the reference voltage, They are respectively a preset proportional coefficient, a preset integral coefficient, and a preset differential coefficient.

[0079] The present invention discloses a photovoltaic power prediction network CH-BiLSTM, which is used to accurately predict photovoltaic power. Traditional attention mechanisms usually assign continuous weights to all time steps, but in photovoltaic power generation prediction, the importance of different time steps varies significantly, and information at some key moments has a decisive influence on the prediction results. To address this problem, the present invention designs a hierarchical attention mechanism within the photovoltaic power prediction network CH-BiLSTM. The hierarchical attention mechanism can more clearly identify and retain key moments through discrete selection operations. At the same time, the hierarchical attention mechanism only retains the most representative The information weight distribution is clearer at each time step, and the interference of redundant information and noise interference is effectively reduced, thereby improving the robustness of the model;

[0080] In addition, after extracting local features, the hierarchical attention mechanism in the present invention re-screens key local information and combines it with global temporal information, enabling the model to more comprehensively capture the dynamic changes of the data.

[0081] The present invention closely combines the hierarchical attention mechanism with a one-dimensional convolutional neural network and a pooling layer. Through discrete screening after local feature extraction, it directly captures the key time steps that have a decisive influence on photovoltaic power prediction, thereby improving the responsiveness and prediction accuracy of the photovoltaic power prediction network CH-BiLSTM to the dynamic changes of photovoltaic data.

[0082] In addition, the present invention collaboratively controls the power grid through the automatic power generation control system AGC and the automatic voltage control system AVC. This collaborative control mechanism breaks the situation where traditional prediction models and power supply system scheduling are independent of each other, and realizes closed-loop management from power generation power prediction to real-time grid regulation, effectively improving the power supply system's ability to cope with fluctuations in photovoltaic power generation output and ensuring the safe and stable operation of the power supply system.

[0083] Referring to it, another aspect of the present invention further provides a power grid control system based on deep learning and photovoltaic power prediction, comprising a power grid, which is controlled by the above-mentioned power grid control method.

[0084] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the ability of ordinary technicians in this field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A grid control method based on deep learning and photovoltaic power prediction, characterized in that: The steps include: S1. Acquire and preprocess photovoltaic data; construct a photovoltaic power prediction network CH-BiLSTM, which includes a sequentially connected one-dimensional convolutional neural network, a pooling layer, a hierarchical attention mechanism, a bidirectional long short-term memory network, and a fully connected layer; S2. Train the photovoltaic power prediction network CH-BiLSTM to obtain the trained photovoltaic power prediction network CH-BiLSTM and deploy it on the device side; use the photovoltaic power prediction network CH-BiLSTM on the device side to perform actual photovoltaic power prediction and obtain the photovoltaic power prediction value. ; S3, automatic power generation control system AGC based on photovoltaic power generation power prediction value , current grid load data and output data of other power generation equipment Calculate the predicted value of photovoltaic power generation In case of changes, other power generation equipment needs to adjust their output and make adjustments to maintain the active power balance of the grid; S4, automatic voltage control system AVC according to the photovoltaic power generation power forecast value And the voltage data of the power grid real-time monitoring Regulating reactive power , in order to change the reactive power distribution in the power grid and stabilize the grid voltage within the allowable range; The S2 specifically includes the following steps: S21, inputting the preprocessed photovoltaic data into a one-dimensional convolutional neural network to obtain multiple local features, and then inputting the multiple local features into a pooling layer to obtain a feature sequence; S22, the hierarchical attention mechanism uses discrete selection operations to retain the previous time steps and convert them into binary indicator variables respectively. The binary indicator variables of time steps are concatenated in time order to obtain weighted feature representation; S23. Use a bidirectional long short-term memory network to perform temporal modeling on the weighted feature representation to obtain a forward hidden state and a backward hidden state. The forward hidden state and the backward hidden state are then concatenated to obtain a final hidden state. The final bidirectional hidden state is linearly mapped through a fully connected layer to generate a predicted value of photovoltaic power. S24. Construct a loss function based on the predicted value and the true value of the photovoltaic power, iterate S21 to S23, minimize the loss function, and adjust the weight of the photovoltaic power prediction network CH-BiLSTM during the iteration process to obtain the trained photovoltaic power prediction network CH-BiLSTM. S25. Deploy the trained photovoltaic power prediction network CH-BiLSTM to the device side, and use the photovoltaic power prediction network CH-BiLSTM on the device side to predict the actual photovoltaic power and obtain the photovoltaic power prediction value. .

2. The grid control method based on deep learning and photovoltaic power prediction according to claim 1 is characterized in that: A dynamic feature fusion strategy is introduced into the one-dimensional convolutional neural network, that is, by adaptively adjusting the feature weights in each time window, the features output by the one-dimensional convolutional neural network can capture local changes; The bidirectional long short-term memory network includes two layers of long short-term memory networks (LSTMs), namely a forward long short-term memory network (LSTM) and a backward long short-term memory network (LSTM). The forward long short-term memory network (LSTM) processes the weighted features obtained by the hierarchical attention mechanism from beginning to end, while the backward long short-term memory network (LSTM) processes the weighted features obtained by the hierarchical attention mechanism from end to beginning, so as to capture more contextual relationships from past and future inputs.

3. The grid control method based on deep learning and photovoltaic power prediction according to claim 1 is characterized in that: The S21 specifically includes the following steps: S211, input the pre-processed photovoltaic data into a one-dimensional convolutional neural network, and the one-dimensional convolutional neural network uses a dynamic feature fusion strategy to extract local features of the pre-processed photovoltaic data to obtain multiple local features. , the calculation formula is: ； in, is the weight parameter of the convolution kernel at the jth position, b is the bias, is the activation function, is the convolution kernel size; Represents the irradiance and power of the time series data at time t+j; S212, multiple local features Input into the pooling layer to obtain the feature sequence H, which is: ; in, Represent the local features of the 1st, 2nd, and Tth time steps respectively.

4. The grid control method based on deep learning and photovoltaic power prediction according to claim 3 is characterized in that: The S22 specifically includes the following steps: S221. Input the feature sequence H into the hierarchical attention mechanism and calculate the score of each time step based on the feature sequence H. The calculation formula is as follows: ； in, represents the score of the i-th time step; 、 Two learnable parameters for score calculation, is a trainable weight vector; Represents the feature vector of the i-th time step; S222, use discrete selection operation to convert the score of each time step into a binary indicator variable ; S223. Preset a critical time step number , for all time steps according to Sort by size, with the highest score binary indicator variable corresponding to each time step is assigned a value of 1, and the other time steps are assigned a value of 0 to retain the one with the highest score time steps; in the above, the number of critical time steps The calculation formula is as follows: ； Where n represents the total number of time steps; S224, the highest retention score The binary indicator variable corresponding to each time step Cascade in chronological order to form the weighted feature representation Z, which is calculated as follows: ； Among them, Concat(·) represents the feature concatenation operation.

5. The grid control method based on deep learning and photovoltaic power prediction according to claim 4 is characterized in that: The S23 specifically includes the following steps: S231. Input the weighted feature representation Z into the forward long short-term memory network LSTM in the bidirectional long short-term memory network. The forward long short-term memory network LSTM processes the weighted feature representation Z from the beginning to the end, and calculates the forward hidden state during the processing. The forward hidden state is calculated as follows: ； in, represents the forward hidden state; is the weighted feature output from the attention mechanism layer at the tth time step, is the forward hidden state at the t-1th time step, and is the forward cell state at the t-1th time step; Represents the forward long short-term memory network LSTM; S232. The backward long short-term memory network LSTM in the bidirectional long short-term memory network processes the weighted feature representation Z from the end to the beginning, and calculates the backward hidden state during the processing; the calculation formula of the backward hidden state is: ； in, Represents the backward hidden state; is the backward hidden state at the t+1th time step, is the backward cell state at the t+1th time step; Represents the backward long short-term memory network LSTM; S233, concatenate the forward hidden state and the backward hidden state to obtain the final bidirectional hidden state; the calculation formula is: ； in, represents the final bidirectional hidden state; S234. Input the final bidirectional hidden state into the fully connected layer, and after linear mapping of the fully connected layer, finally generate the predicted value of photovoltaic power.

6. The grid control method based on deep learning and photovoltaic power prediction according to claim 5, characterized in that: The S3 specifically includes the following steps: S31, the photovoltaic power generation power prediction value , current grid load data and output data of other power generation equipment Transmitted together to the automatic power generation control system AGC; S32, calculate the photovoltaic power generation power prediction value based on the power balance principle of the automatic power generation control system AGC In case of changes, the output of other power generation equipment needs to be adjusted , as follows: ； in, is the photovoltaic power generation power at the previous moment, is the load power at the previous moment; S33, the automatic power generation control system AGC adjusts the output of other power generation equipment according to the calculation Send adjustment instructions to adjust power generation output, maintain the active power balance of the power grid, and ensure that the power grid frequency is stable within the specified range.

7. The grid control method based on deep learning and photovoltaic power prediction according to claim 6, characterized in that: The S4 specifically includes the following steps: S41, automatic voltage control system AVC according to the photovoltaic power generation power forecast value And the voltage data of the power grid real-time monitoring Regulating reactive power ; S42, automatic voltage control system AVC based on reactive power Adjust the reactive power compensation equipment or adjust the excitation current of the generator to change the reactive power distribution in the power grid and stabilize the grid voltage within the allowable range.

8. The grid control method based on deep learning and photovoltaic power prediction according to claim 7, characterized in that: The reactive power in S41 The calculation formula is as follows: ； in, is the reference voltage, They are respectively a preset proportional coefficient, a preset integral coefficient, and a preset differential coefficient.

9. A power grid control system based on deep learning and photovoltaic power prediction, comprising a power grid, characterized in that: The power grid is regulated by the power grid regulation method according to any one of claims 1 to 8.

Images