Distribution network new energy output prediction model based on artificial intelligence

Abstract

The invention belongs to the technical field of power systems, and particularly relates to a distribution network new energy output prediction model based on artificial intelligence, and the model comprises an SL-iGATv2 space-time prediction model architecture, and a distribution network new energy output prediction method of the SL-iGATv2 space-time prediction model architecture comprises the following steps: S1, collecting original data; comprise historical generating capacity, weather data, season information, geographic position parameters and equipment performance indexes; s2, a GATv2 graph neural encoder is built, after collected original data are input, initialization and normalization processing of data information are carried out, feature extraction and fusion are carried out on different data sources, and then final feature output of each node is carried out; s3, a Sat iotempora l LSTM (Long Short Term Memory) model is built; and S4, a Softmax layer is established at the tail end of the Sat iotempora l LSTM model. The method integrates use space, history and policy information, has high adaptability and generalization ability, can cope with variable meteorological conditions and user behaviors, and keeps strong robustness and fault tolerance at the same time.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a distribution network renewable energy output prediction model based on artificial intelligence. Background Art

[0002] With the continuous growth of renewable energy in the power system, the prediction and dispatch optimization of renewable energy output has become an important research field. The accuracy of renewable energy output prediction and the efficiency of dispatch optimization have an important impact on the operation and stability of the power system. Specifically, renewable energy output prediction refers to the prediction of the output of renewable energy power generation devices such as wind power and solar power, and the subsequent power system dispatch based on the prediction results.

[0003] At present, the commonly used methods for predicting the output of new energy in distribution networks are recurrent neural networks (RNN), long short-term memory networks (LSTM), ensemble learning methods (Gradient Boosting Tree and XGBoost), statistical methods, etc. However, they are basically predictions based on historical data, with less consideration of weather factors, life and economic policy factors, and cannot form a unified prediction system. Moreover, the use of geographic location relationship quantity is relatively incomplete. Therefore, exploring a spatiotemporal prediction model with spatial information combined with historical data and adapted to new energy output is the key technology for this solution to solve the prediction accuracy. To this end, we provide a distribution network new energy output prediction model based on artificial intelligence. Summary of the invention

[0004] In response to the above-mentioned technical problems, the present invention proposes a SL-i GATv2 spatiotemporal prediction neural network model that can integrate usage space, history, and policy information, improves usage functions, and fully guarantees the integrity of usage functions. It is an artificial intelligence-based distribution network new energy output prediction model.

[0005] In order to achieve the above-mentioned object, the technical solution adopted by the present invention is a distribution network new energy output prediction model based on artificial intelligence, including an SL-i GATv2 spatiotemporal prediction model architecture, and the method for predicting the distribution network new energy output of the SL-i GATv2 spatiotemporal prediction model architecture includes the following steps:

[0006] S1: First, collect raw data, including historical power generation, weather data, seasonal information, geographic location parameters, and equipment performance indicators;

[0007] S2: Build the GATv2 graph neural encoder, input the collected raw data, initialize and normalize the data information, extract and fuse features from different data sources, and then perform the final output features of each node;

[0008] S3: Build a Spatial Temporal LSTM model to model spatiotemporal sequence data to capture long-term dependencies and spatiotemporal dynamics in time series data;

[0009] S4: A Softmax layer is built at the end of the Spat iotempora l LSTM model to realize the weight allocation of different influencing factors, so as to measure the influence of different factors on the prediction of new energy output of distribution network, and the prediction result of various factors is obtained by multiplying the output of Softmax layer with the prediction result of Spat iotempora l LSTM model.

[0010] Preferably, step 2 comprises the following steps:

[0011] S2.1: In information initialization, for a target node u and one of its neighbor nodes v, a function with shared parameters is used to calculate their attention or compatibility. The formula for calculating the attention or compatibility of the target node and one of its neighbor nodes is:

[0012] e_{uv}=LeakyReLU(W^T[h_u||h_v])

[0013] Among them, h_u and h_v are the feature vectors of nodes u and v, [h_u||h_v] represents the concatenation of feature vectors, W^T is the weight matrix of the attention mechanism, and LeakyReLU is a nonlinear activation function.

[0014] S2.2: Use the softmax function to perform normalization processing to obtain a probability distribution, and the formula for obtaining the probability distribution is: α_{uv}=softmax(e_{uv});

[0015] S2.3: Use the calculated attention coefficient to perform weighted summation on the features of neighbor nodes to aggregate information. The formula for aggregating information is: h'_u = ∑_{v∈N(u)}α_{uv}Wh_v, where N(u) represents the neighbor set of node u and W is another transformation matrix;

[0016] S2.4: After obtaining the normalized attention coefficient, the linear combination of the features corresponding to the normalized attention coefficient is calculated as the final output feature of each node. The formula of the final output feature is:

[0017]

[0018] Among them, σ represents the si gmo id activation function, W and h represent the learnable parameters in training;

[0019] S2.5: Build an infect on GATv2 model to dynamically adjust the behavior of the model, so as to dynamically adjust the attention weights according to the relationship and importance between nodes, so as to capture the local features and global structure of the graph data.

[0020] Preferably, the calculation formula of the attention system of the attention mechanism in step S2.3 is expressed as:

[0021]

[0022] Among them, || represents a concatenation operation, LeakyReLU represents an activation function, and Wh represents a learnable parameter.

[0023] Preferably, the formula for dynamically adjusting the attention weight in step S2.5 is expressed as:

[0024]

[0025] Preferably, a gating mechanism is built in the Spatial LSTM model of step S3, wherein the gating mechanism includes an input gate, a forget gate, a cell state update and an output gate, and includes the following steps:

[0026] S3.1: Use the input gate to view the current input data and the hidden state of the previous time step, and determine the information to be stored. The formula for determining the information to be stored by the input gate is:

[0027] i t =σ(W xi ·[x t ;h t-1 ]+b i )

[0028] Among them, x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xi is the weight matrix corresponding to the input gate, b i is the bias term, σ represents the si gmo id activation function;

[0029] S3.2: Use the forget gate to determine the unimportant information that is allowed to be discarded in the model. The formula of the forget gate that allows the model to discard information is: t =σ(W xf ·[x t ;h t-1 ]+b f ), where x t is the input of the current time step, h t-1 is the hidden state of the previous time step, Wxf is the weight matrix of the forget gate, bf is the bias term;

[0030] S3.3: According to the signal input in S3.1 and S3.2, the cell state is updated. The formula for updating the cell state is:

[0031] c t =f t ·c t-1 +i t tanh(W xc ·[x t ;h t-1 ]+b c )

[0032] Among them, c t is the cell state at the current time step, c t-1 is the cell state at the previous time step, f t is the output of the forget gate, i t is the output of the input gate, W xc is the associated weight matrix, b c is the bias term;

[0033] S3.4: Use the output gate to determine how much information is output in the cell state update to the final result. The formula for determining how much information is output by the output gate is:

[0034] o t =σ(W xo ·[x t ;h t-1 ]+b o )

[0035] Among them, W xo is the weight matrix of the output gate, b o is the bias term.

[0036] Preferably, step S4 comprises the following steps:

[0037] S4.1: Assuming that there are n influencing factors, a probability distribution is generated for them using the constructed Softmax layer. The probability distribution determination formula of the Softmax layer is:

[0038] P=(p1,p2,...,p n )

[0039] Among them, p i Represents the weight of the ith factor.

[0040] S4.2: Assume that the LSTM layer makes a prediction based on each factor, denoted as:

[0041] L=(l1,l2,...,l n );

[0042] S4.3: Use the output in step S4.1 as a weight to perform weighted summation on the predicted values ​​in S4.2. The formula for determining the weighted prediction result is expressed as: Among them, p i ·l i is the weighted estimate of the (i)th factor.

[0043] Compared with the prior art, the advantages and positive effects of the present invention are:

[0044] 1. The present invention provides an artificial intelligence-based distribution network renewable energy output prediction model, which combines the advantages of the GATv2 graph neural encoder in processing non-Euclidean data and the powerful ability of LSTM in capturing long-term dependencies of time series; the SL-i GATv2 spatiotemporal prediction model obtained by this combination can effectively extract key features from complex spatiotemporal data and accurately predict the generation and consumption trends of renewable energy. This method has high adaptability and generalization capabilities, can cope with changing meteorological conditions and user behaviors, and maintain strong robustness and fault tolerance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 This is the architecture diagram of the SL-i GATv2 model provided by the present invention;

[0047] Figure 2 A network architecture diagram of the Spatial Temporal LSTM model provided by the present invention; DETAILED DESCRIPTION

[0048] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0050] Examples, such as Figure 1 As shown, a distribution network new energy output prediction model based on artificial intelligence, wherein the essence of the new energy output prediction model is the knowledge modeling expression of mapping exogenous information to new energy output information. Based on this, the advantages of GATv2 graph neural encoder in processing non-Euclidean data and the powerful ability of LSTM in capturing long-term dependencies of time series are integrated. The SL-i GATv2 spatiotemporal prediction model architecture obtained under this combination, further said: The method for predicting the new energy output of the distribution network of the SL-iGATv2 spatiotemporal prediction model architecture includes the following steps:

[0051] S1: First, collect raw data, including historical power generation, weather data, seasonal information, geographic location parameters, and equipment performance indicators;

[0052] Specifically described as follows: This step is data processing and feature engineering, which is responsible for collecting raw data, including but not limited to historical power generation, weather data (such as temperature, wind speed, sunshine, etc.), seasonal information, geographic location parameters, and equipment performance indicators; the data needs to be cleaned, standardized (i.e. initialized) and normalized to facilitate model processing, and then provide convenient conditions for subsequent feature extraction and fusion of different data sources. For example, weather and seasonal factors can be used to identify patterns through historical data analysis, geographic factors may include information such as the latitude and longitude of the site, altitude, etc., and equipment performance may include equipment model, efficiency, maintenance records, etc.;

[0053] S2: Build the GATv2 graph neural encoder, input the collected raw data, initialize and normalize the data information, extract and fuse features from different data sources, and then perform the final output features of each node;

[0054] The specific description is as follows: The GATv2 graph neural encoder is a neural network model based on graph structure data. It introduces an attention mechanism to weight the neighbor node features of a node, thereby learning node representation. This model is particularly suitable for processing structured data, such as social networks, molecular structures, and transportation networks;

[0055] S3: Build a Spatial Temporal LSTM model to model spatiotemporal sequence data to capture long-term dependencies and spatiotemporal dynamics in time series data;

[0056] The specific description is as follows: The Spatial Temporal LSTM model is one of the core parts of the SL-iGATv2 spatiotemporal prediction model architecture. This part includes multiple Spatial Temporal LSTM units, each of which is responsible for modeling spatiotemporal series data, and can capture the long-term dependencies and spatiotemporal dynamics in the time series data, and is suitable for processing data involving spatiotemporal changes;

[0057] S4: A Softmax layer is built at the end of the Spat iotempora l LSTM model to realize weight allocation of different influencing factors, so as to measure the influence of different factors on the prediction of the output of new energy in the distribution network, and the prediction result of the Spat iotempora l LSTM model is multiplied by the output of the Softmax layer to obtain the prediction result under various factors.

[0058] Specifically described as follows: The role of the SoftMax layer after the LSTM network is to further process the sequence features generated by the LSTM layer to output the final weights and prediction results. Adding a Softmax layer at the end of the LSTM model can assign weights to different influencing factors (such as weather, policies, etc.), and these weights reflect the contribution of each factor to the final load prediction; by multiplying the weight output by the Softmax layer with the prediction result of the LSTM, the weighted prediction value can be obtained to more accurately estimate the distribution network load.

[0059] In the above process: through the established SL-iGATv2 spatiotemporal prediction model, the advantages of the GATv2 graph neural encoder in processing non-Euclidean data and the powerful ability of LSTM in capturing the long-term dependencies of time series are integrated, which can effectively extract key features from complex spatiotemporal data and accurately predict the generation and consumption trends of new energy. This method has high adaptability and generalization capabilities, can cope with changeable meteorological conditions and user behaviors, and at the same time, maintain strong robustness and fault tolerance; the SL-i GATv2 spatiotemporal prediction neural network model of the present invention integrates the use of space, history, and policy information to improve the use function and fully guarantee the integrity of the use function.

[0060] In order to effectively obtain the factor position, quantity relationship and weight relationship of the input data, in the scenario of predicting the power generation of renewable energy, the nodes represent different renewable energy power plants, weather stations or geographical areas, and a certain relationship between the nodes is established through the edges. Based on the historical power generation data, weather data and geographical location, a feature vector is assigned to each node for information initialization. Further, step 2 includes the following steps:

[0061] S2.1: In information initialization, a function with shared parameters is used to calculate the attention or compatibility of the target node u and one of its neighbor nodes v. The formula for calculating the attention or compatibility of the target node and one of its neighbor nodes is:

[0062] e_{uv}=LeakyReLU(W^T[h_u||h_v])

[0063] Among them, h_u and h_v are the feature vectors of nodes u and v, [h_u||h_v] represents the concatenation of feature vectors, W^T is the weight matrix of the attention mechanism, and LeakyReLU is a nonlinear activation function.

[0064] S2.2: Use the softmax function to perform normalization to obtain a probability distribution. The formula for obtaining the probability distribution is: α_{uv}=softmax(e_{uv});

[0065] S2.3: Use the calculated attention coefficient to perform weighted summation on the features of neighboring nodes to aggregate information. The formula for aggregating information is: h'_u = ∑_{v∈N(u)}α_{uv}Wh_v, where N(u) represents the neighbor set of node u and W is another transformation matrix;

[0066] S2.4: After obtaining the normalized attention coefficient, the linear combination of the corresponding features is calculated using the normalized attention coefficient as the final output feature of each node. The formula for the final output feature is:

[0067]

[0068] Among them, σ represents the si gmo id activation function, W and h represent the learnable parameters in training; after enough layers, each node will have a rich embedding representation that captures the interactions between the node and its neighbors.

[0069] S2.5: Build an infect on GATv2 model to dynamically adjust the behavior of the model, so as to dynamically adjust the attention weights according to the relationship and importance between nodes, so as to capture the local features and global structure of the graph data;

[0070] The specific description is: Infection GATv2 is one of the core parts of the model. In the infection model, the propagation path and direction are often affected by external factors (such as wind direction). SL-i GATv2 uses this idea to dynamically adjust the behavior of the model, realize information transmission and relationship modeling between nodes in graph data, and can dynamically adjust the attention weight according to the relationship and importance between nodes, better capture the local characteristics and global structure of graph data, and help improve the model's spatial dependency.

[0071] To further increase the depth and expressiveness of the model, the prediction model stacks multiple such attention layers. Each layer updates the feature representation of the node. The attention layer is a single-layer feedforward neural network composed of a weight vector Parameterized, and LeakyReLU nonlinear activation function is applied (negative input slope α = 0.2). After full expansion, the calculation formula of the attention system of the attention mechanism in step S2.3 is expressed as:

[0072]

[0073] Among them, || represents a concatenation operation, LeakyReLU represents an activation function, and Wh represents a learnable parameter.

[0074] In order to ensure the effectiveness of the infect ion GATv2 model, the formula for dynamically adjusting the attention weight in step S2.5 is expressed as:

[0075]

[0076] Further: In this way, GATv2 can effectively learn the representation of nodes in the graph, while retaining the important information of the original features and being able to adapt to data of various graph structures. In the forecast of renewable energy power generation, this helps the model better understand and predict how different factors (such as weather, season, geographic location, and equipment performance) interact to affect power generation.

[0077] In order to fully guarantee the specific extraction of spatiotemporal features by the Spatial Temporal LSTM model, a gating mechanism is built in the Spatial Temporal LSTM model in step S3. The gating mechanism includes an input gate, a forget gate, a cell state update and an output gate, and includes the following steps:

[0078] S3.1: Use the input gate to view the current input data and the hidden state of the previous time step to determine the information to be stored. The formula for determining the information to be stored by the input gate is:

[0079] i t =σ(Wxi ·[x t ;h t-1 ]+b i )

[0080] Among them, x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xi is the weight matrix corresponding to the input gate, b i is the bias term, σ represents the si gmo id activation function;

[0081] S3.2: Use the forget gate to determine the unimportant information that is allowed to be discarded in the model. The formula for the forget gate to allow the model to discard information is: t =σ(W xf ·[x t ;h t-1 ]+b f ), where x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xf is the weight matrix of the forget gate, bf is the bias term;

[0082] S3.3: Update the cell state according to the signal input in S3.1 and S3.2. The formula for updating the cell state is:

[0083] c t =f t ·c t-1 +i t tanh(W xc ·[x t ;h t-1 ]+b c )

[0084] Among them, c t is the cell state at the current time step, c t-1 is the cell state at the previous time step, f t is the output of the forget gate, i t is the output of the input gate, W xc is the associated weight matrix, b c is the bias term;

[0085] S3.4: Use the output gate to determine how much information is output in the cell state update to the final result. The formula for determining how much information is output by the output gate is:

[0086] o t =σ(W xo ·[x t ;h t-1 ]+bo )

[0087] Among them, W xo is the weight matrix of the output gate, b o is the bias term;

[0088] The specific description is: Figure 2 As shown in the figure, Spat iotempora l LSTM adopts a hierarchical structure, in which each layer extracts specific spatiotemporal features. This structure allows the network to capture different levels of abstract information at different levels, and this information can be further utilized and refined in subsequent layers. Key information is retained by updating the temporal and spatial states. The temporal state is responsible for maintaining the continuity of the time series, while the spatial state is responsible for recording changes in spatial information. This separate state update mechanism enables the network to process complex spatiotemporal dependencies more effectively. The parallel structure design of Spat iotempora lLSTM enables the model to process information independently between different layers. Although this design may lead to insufficient information transfer between layers, it also allows the model to focus on specific types of feature extraction at different levels, thereby improving the model's ability to process complex spatiotemporal data.

[0089] The node connections and weights output by the graph neural network (GNN) can be used as features to input into the Spat iotempora lLSTM model. This information reflects the interdependence and spatial correlation between different energy sites. Using the structured data provided by the graph neural network, time series data can be constructed, which contains chronological information such as historical power generation, weather conditions, and geographical factors. In LSTM, the input gate determines what information will be stored in the cell state. This gating mechanism is determined by looking at the current input data and the hidden state of the previous time step. The forget gate is responsible for determining which information should be discarded from the cell state. This allows the model to discard information that is no longer important, allowing the model to better handle long-term dependencies. Based on the signals of the input gate and the forget gate, the cell state is updated. This state is a core feature of LSTM, which can maintain important information in long sequences. The output gate determines how much information is output from the cell state to the final result. In this way, the model can generate a prediction result for the next time step. Through the above-mentioned gating mechanism, Spatial Temporal LSTM can not only capture the dependency of time series data, but also combine the spatial characteristics provided by the graph neural network to more comprehensively understand and predict the changes in renewable energy power generation. The model is trained using historical data, and the prediction error is minimized by continuously adjusting the weights. During the training process, the network parameters can be optimized through the back-propagation algorithm to obtain more accurate prediction results. After sufficient training, the model will be able to predict future renewable energy power generation, taking into account multiple factors such as historical data, weather changes, seasonal effects, geographical environment, and equipment performance.

[0090] In order to output the final weights and prediction results, the role of the SoftMax layer after the LSTM network is to further process the sequence features generated by the LSTM layer to output the final weights and prediction results; further: adding a Softmax layer at the end of the LSTM model can assign weights to different influencing factors (such as weather, policies, etc.), and these weights reflect the contribution of each factor to the final load prediction; by multiplying the weight output by the Softmax layer with the prediction result of the LSTM, we can get the weighted prediction value and estimate the distribution network load more accurately. Step S4 includes the following steps:

[0091] S4.1: Assuming that there are n influencing factors, a probability distribution is generated for them using the constructed Softmax layer. The probability distribution determination formula of the Softmax layer is:

[0092] P=(p1,p2,...,p n )

[0093] Among them, p iRepresents the weight of the ith factor.

[0094] S4.2: Assume that the LSTM layer makes a prediction based on each factor, denoted as:

[0095] L=(l1,l2,...,l n );

[0096] S4.3: Use the output in step S4.1 as a weight to perform weighted summation on the predicted values ​​in S4.2. The formula for determining the weighted prediction result is expressed as: Among them, p i ·l i is the weighted estimate of the (i)th factor;

[0097] Specifically described as follows: Steps S4.2 and S4.3 are weighted prediction results. The role of the Softmax layer is to learn a probability distribution that can measure the degree of influence of different factors on load forecasting. For example, if the probability of weather factors is high, it means that weather has a greater impact on load forecasting. By multiplying the output of the Softmax layer with the prediction result of LSTM, we get a prediction result that comprehensively considers the importance of various factors. This processing method allows the model to more flexibly adjust the proportion of each factor in the final prediction.

[0098] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A distribution network renewable energy output prediction model based on artificial intelligence, characterized in that: The method of predicting the output of new energy in a distribution network using the SL-iGATv2 spatiotemporal prediction model architecture comprises the following steps: S1: First, collect raw data, including historical power generation, weather data, seasonal information, geographic location parameters, and equipment performance indicators; S2: Build the GATv2 graph neural encoder, input the collected raw data, initialize and normalize the data information, extract and fuse features from different data sources, and then perform the final output features of each node; S3: Build a Spatiotemporal LSTM model to model spatiotemporal sequence data to capture long-term dependencies and spatiotemporal dynamics in time series data; S4: A Softmax layer is built at the end of the Spatiotemporal LSTM model to realize the weight allocation of different influencing factors, so as to measure the influence of different factors on the prediction of renewable energy output of distribution network, and the prediction result of various factors is obtained by multiplying the output of Softmax layer with the prediction result of Spatiotemporal LSTM model.

2. The artificial intelligence-based distribution network new energy output prediction model according to claim 1 is characterized in that: The step 2 comprises the following steps: S2.1: In information initialization, for a target node u and one of its neighbor nodes v, a function with shared parameters is used to calculate their attention or compatibility. The formula for calculating the attention or compatibility of the target node and one of its neighbor nodes is: e_{uv}=LeakyReLU(W^T[h_u||h_v]) Among them, h_u and h_v are the feature vectors of nodes u and v, [h_u||h_v] represents the concatenation of feature vectors, W^T is the weight matrix of the attention mechanism, and LeakyReLU is a nonlinear activation function. S2.2: Use the softmax function to perform normalization processing to obtain a probability distribution, and the formula for obtaining the probability distribution is: α_{uv}=softmax(e_{uv}); S2.3: Use the calculated attention coefficient to perform weighted summation on the features of neighbor nodes to aggregate information. The formula for aggregating information is: h'_u = ∑_{v∈N(u)}α_{uv}Wh_v, where N(u) represents the neighbor set of node u and W is another transformation matrix; S2.4: After obtaining the normalized attention coefficient, the linear combination of the features corresponding to the normalized attention coefficient is calculated as the final output feature of each node. The formula of the final output feature is: Among them, σ represents the sigmoid activation function, W and h represent the learnable parameters in training; S2.5: Build the infection GATv2 model to dynamically adjust the behavior of the model to dynamically adjust the attention weights according to the relationship and importance between nodes, so as to capture the local features and global structure of the graph data.

3. The artificial intelligence-based distribution network new energy output prediction model according to claim 2 is characterized in that: The calculation formula of the attention system of the attention mechanism in step S2.3 is expressed as: Among them, || represents a concatenation operation, LeakyReLU represents an activation function, and Wh represents a learnable parameter.

4. The artificial intelligence-based distribution network new energy output prediction model according to claim 3 is characterized in that: The formula for dynamically adjusting the attention weight in step S2.5 is expressed as:

5. The artificial intelligence-based distribution network new energy output prediction model according to claim 4 is characterized in that: A gating mechanism is built in the Spatiotemporal LSTM model of step S3, wherein the gating mechanism includes an input gate, a forget gate, a cell state update and an output gate, and includes the following steps: S3.1: Use the input gate to view the current input data and the hidden state of the previous time step, and determine the information to be stored. The formula for determining the information to be stored by the input gate is: i t =σ(W xi ·[x t ;h t-1 ]+b i ) Among them, x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xi is the weight matrix corresponding to the input gate, b i is the bias term, σ represents the sigmoid activation function; S3.2: Use the forget gate to determine the unimportant information that is allowed to be discarded in the model. The formula of the forget gate that allows the model to discard information is: t =σ(W xf ·[x t ;h t-1 ]+b f ), where x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xf is the weight matrix of the forget gate, bf is the bias term; S3.3: According to the signal input in S3.1 and S3.2, the cell state is updated. The formula for updating the cell state is: c t =f t ·c t-1 +i t ·tanh(W xc ·[x t ;h t-1 ]+b c ) Among them, c t is the cell state at the current time step, c t-1 is the cell state at the previous time step, f t is the output of the forget gate, i t is the output of the input gate, W xc is the associated weight matrix, b c is the bias term; S3.4: Use the output gate to determine how much information is output in the cell state update to the final result. The formula for determining how much information is output by the output gate is: the t =σ(W xo ·[x t ;h t-1 ]+b o ) Among them, W xo is the weight matrix of the output gate, b o is the bias term.

6. The artificial intelligence-based distribution network new energy output prediction model according to claim 5, characterized in that: The step S4 comprises the following steps: S4.1: Assuming that there are n influencing factors, a probability distribution is generated for them using the constructed Softmax layer. The probability distribution determination formula of the Softmax layer is: P=(p1,p2,...,p n ) Among them, p i Represents the weight of the ith factor. S4.2: Assume that the LSTM layer makes a prediction based on each factor, denoted as: L＝(l1,l2,...,l n )； S4.3: Use the output in step S4.1 as a weight to perform weighted summation on the predicted values ​​in S4.

2. The formula for determining the weighted prediction result is expressed as: Among them, p i ·l i is the weighted estimate of the (i)th factor.

Images