A source-load joint prediction method and system based on ring-coupled multi-expert neural network

Abstract

The present application relates to the technical fields of power system source load prediction, new energy power prediction, load prediction, multivariate time series modeling and deep learning prediction, and discloses a source load joint prediction method and system based on ring coupling multi-expert neural network, which takes multiple neural network sub-models with different modeling responsibilities as prediction nodes in a ring topology, each node is responsible for a specific information domain or a prediction sub-task, and performs multiple rounds of state exchange and residual correction through a ring message passing mechanism, and finally outputs the joint prediction results of source side output, load side demand and net load. Through ring message passing, the present application realizes step-by-step collaborative reasoning, so that the source side prediction error, load side peak-valley change, weather mutation and net load risk can be circulated and corrected among multiple nodes; and the complex coupling relationship between the source side, the load side, the weather, the time, the operation state and the event risk can be better expressed.

Description

Technical Field

[0001] This invention relates to the fields of power system source-load prediction, new energy power prediction, load prediction, multivariate time series modeling and deep learning prediction technology, and in particular to a source-load joint prediction method and system based on a ring-coupled multi-expert neural network. Background Technology

[0002] With the increasing penetration rate of renewable energy, net load forecasting and combined source-load forecasting have become crucial foundations for reliable power system operation, optimal resource allocation, and economic dispatch. Existing research indicates that in scenarios with a high proportion of renewable energy integration, net load forecasting plays a key role in balancing supply and demand, improving resource utilization efficiency, and ensuring reliable power system operation. Meanwhile, deep learning research in the field of short-term load forecasting is shifting from simply pursuing forecast accuracy to focusing on online adaptability, robustness, and engineering deployability.

[0003] Based on the retrieval analysis, existing source load prediction and multivariate time series prediction technologies can be roughly divided into the following categories.

[0004] The first category is prediction methods based on a single deep neural network backbone. These methods typically employ structures such as recurrent neural networks, convolutional neural networks, Transformers, temporal convolutional networks, or multilayer perceptrons. They concatenate historical power, load, meteorological, and temporal features and input them into a unified model, directly outputting future predictions. While these methods are simple to implement and have a clear training process, they usually rely on a single backbone network to complete the entire modeling task. This can easily lead to feature aliasing between heterogeneous information such as source-side output, load-side demand, meteorological disturbances, and rhythmic cycles, making it difficult to reflect the specialized modeling capabilities of different information sources.

[0005] The second category comprises time series forecasting methods based on decomposition, periodic modeling, and block modeling. Autoformer improves trend-period modeling capabilities in long-term forecasting through sequence decomposition and autocorrelation mechanisms; TimesNet enhances the expressive power of multi-period time series patterns by unfolding complex temporal variations into intra-period and inter-period variations; and PatchTST reduces attention computational overhead while preserving local semantic information through time fragmentation and channel-independent modeling. These methods demonstrate that specialized design targeting the internal structure of time series can significantly improve forecasting capabilities. However, they primarily address the problem of time dependency modeling within a single or multivariate sequence, and have not yet fully addressed the closed-loop collaborative reasoning problem among multiple forecasting subsystems such as the source side, load side, meteorological side, and operational side.

[0006] The third category is multi-task learning and multi-energy load joint forecasting methods. Existing research on multi-energy load forecasting transforms the forecasting of various loads such as electricity, heat, and cooling into a hierarchical multi-task learning problem, and combines spatiotemporal attention mechanisms and gated temporal convolutional networks to handle the coupling relationships between different energy loads. This type of method can utilize the correlation between tasks at the shared representation level, but most still use shared encoders, task branching, or attention fusion structures. The information flow between tasks is usually one-off or hierarchical, lacking multi-round negotiation, state backflow, and cycle consistency constraints between task nodes.

[0007] The fourth category comprises prediction methods based on expert blending, base models, and cue learning. Methods like Time-MoE introduce the concept of expert blending into time series prediction, enabling different experts to handle different patterns. Time-LLM, on the other hand, reprograms time series segments into a text prototype space and uses a prompt-as-prefix approach to enhance the understanding of time series by large language models. These methods demonstrate that different experts or different semantic spaces can improve the model's generalization and pattern representation capabilities. However, existing expert blending methods typically employ gating selection or sparse activation mechanisms, lacking circular information transfer and closure consistency correction among experts. Cue learning methods, while emphasizing cross-modal semantic alignment, have not yet constructed a closed-loop collaborative structure among multiple prediction subsystems for source-load prediction.

[0008] While currently available source load forecasting and multivariate time series forecasting methods have strong nonlinear modeling capabilities, they still have the following shortcomings.

[0009] First, existing methods mostly employ a single backbone network or a static multi-branch fusion structure, lacking the ability to achieve closed-loop interaction between prediction subsystems. There are significant coupling relationships between source-side output, load-side demand, meteorological disturbances, and operational status, but existing models typically only concatenate multiple types of variables before inputting them into a unified network, or perform feature fusion at the end, making it difficult to express the multi-round influences and feedback correction relationships between different information domains.

[0010] Second, existing multi-task learning methods typically emphasize shared encoding and task branch outputs, but lack explicit information feedback mechanisms between tasks. Source-side prediction errors may affect net load prediction, and load-side peak-valley variations may also alter source-load balance risks. Existing multi-task structures often lack mechanisms to transfer residual information, confidence information, and risk information from one prediction task to another.

[0011] Third, existing expert hybrid methods typically select experts through gating networks, resulting in a lack of structured collaboration among experts. While different experts can learn different patterns, in most cases they are passively selected parallel units rather than collaborative nodes with sequential transmission, state updates, and cycle consistency constraints. Therefore, it is difficult for experts to form interpretable information flow paths.

[0012] Fourth, while existing graph neural network methods can express the spatial or topological relationships between nodes, in source-load prediction scenarios, the source side, load side, meteorological side, time side, and event side do not necessarily correspond to fixed physical topological nodes. If a fixed graph structure is directly adopted, it is easily affected by the quality of graph construction; if fully connected attention is used, it may introduce too many irrelevant interactions, reducing the stability and interpretability of the model.

[0013] Fifth, existing forecasting methods lack sufficient constraints on the net load balance relationship. Source-load forecasting not only requires improving the forecast accuracy of both source-side output and load-side demand, but also ensuring physical consistency between source-side, load-side, and net load forecasts. For example, net load typically has a clear balance relationship with load demand and renewable energy output. If the model only outputs multiple forecast results separately without imposing consistency constraints, it may result in small individual forecast errors but unstable overall net load results.

[0014] Sixth, existing models struggle to explain the information transfer process between different forecast branches. For power dispatching applications, models not only need to output future numerical results but also explain whether the current forecast is primarily driven by meteorological disturbances, load cycles, source-side ramp-up, or event risks. Traditional black-box deep models typically fail to provide this structured explanation.

[0015] In summary, while existing technologies have made progress in long-sequence modeling, multi-scale decomposition, multi-task prediction, expert hybridization, and semantic enhancement, they still share a common shortcoming: most methods treat multi-source information fusion as a one-time splicing, weighting, or gating selection, rather than modeling the source side, load side, meteorological side, time side, operational status side, and event risk side as multiple interactive, feedback-enabled, and cyclically correctable prediction nodes. Therefore, in scenarios with complex source-load coupling relationships, locally concentrated prediction errors, and rapidly changing net loads, a novel source-load prediction network capable of achieving multi-expert closed-loop collaborative reasoning is still needed. Summary of the Invention

[0016] The technical problem to be solved by the present invention is to provide a source-load joint prediction method and system based on a loop-coupled multi-expert neural network that can realize multi-expert closed-loop collaborative reasoning in scenarios with complex source-load coupling relationships, local concentration of prediction errors, and rapid changes in net load.

[0017] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows: a source-load joint prediction method based on a ring-coupled multi-expert neural network is provided. It uses multiple neural network sub-models with different modeling responsibilities as prediction nodes in a ring topology. Each node is responsible for a specific information domain or prediction sub-task. Through a ring-shaped message passing mechanism, multiple rounds of state exchange and residual correction are performed, and finally the joint prediction results of source-side output, load-side demand and net load are output.

[0018] Preferably, it includes the following steps: Step 1: Obtain and construct the historical input sequence for joint source-load prediction; Step 2: Construct ring-coupled multi-expert neural network nodes; Step 3: Each expert node performs local coding; Step 4: Construct a circular message passing mechanism; Step 5: Construct a cyclic gating and conflict suppression mechanism; Step Six: Construct source-load consistency and loop consistency constraints; Step 7: Perform multi-round circular collaborative reasoning and adaptive stopping; Step 8: Output the source-side, load-side, and net load prediction results; Step 9: Construct joint training objectives.

[0019] Preferably, in step one, historical multivariate time series data of the target area or target power station are first obtained; the input data includes, but is not limited to, source-side power output data, load-side demand data, meteorological data, time calendar data, power grid operation status data, and external event data; Source-side output data may include wind power, photovoltaic power, distributed renewable energy output, and energy storage charging and discharging power; load-side demand data may include total load, industrial load, commercial load, residential load, charging pile load, and park load; meteorological data may include wind speed, wind direction, irradiance, temperature, humidity, air pressure, precipitation, and cloud cover; time calendar data may include hour, day type, weekday, holiday, season, and month; operational status data may include line power flow, node voltage, energy storage status, demand response status, and day-ahead plan; the above data are time-aligned according to a unified sampling interval to construct a historical input sequence: , in, Indicates the length of the history window. Indicates time Multivariate inputs; prediction targets include future... Step-by-step power output sequence, load-side demand sequence, and net load sequence: , in, Indicates the power output from the source side. Indicates load-side demand. This represents the net load; the net load is determined by the difference between load-side demand and source-side renewable energy output, namely: .

[0020] Preferably, the source-load forecasting task is divided into several complementary information domain modeling tasks, and an expert neural network node is constructed for each type of task. Multiple expert nodes are connected in a ring topology to form a closed collaborative network. The ring network includes: a source-side output expert node, used to model the historical changes, ramp-up events, and meteorological driving relationships of new energy output such as wind power and photovoltaics; a load-side demand expert node, used to model the periodicity, peak-valley characteristics, holiday effects, and user behavior changes of load demand; a meteorological disturbance expert node, used to model the combined impact of meteorological variables on source-side output and load-side demand; a time rhythm expert node, used to model daily, weekly, seasonal, and holiday patterns; an operational status expert node, used to model the impact of source-grid-load-storage operational status, energy storage charging and discharging, demand response, and scheduling plans on future source-load changes; and an event risk expert node, used to identify high-risk forecast segments such as source-side ramp-up, load surge, meteorological abrupt changes, and sharp changes in net load. In different application scenarios, four, five, six, or more nodes can be set; nodes can also be deleted or added according to the availability of actual data. Assume the ring network has a total of The expert node, the first Each expert node is represented as: , in, Indicates a local encoder. This indicates that the node status is updated in the network. This indicates a circular message generation network. This represents the node prediction decoder.

[0021] Preferably, in step three, the input sequence is divided into different sub-inputs according to the information domain: , in, Indicates source-side input, Indicates load-side input. Indicates meteorological input. This indicates the input of a time calendar. Indicates the running status input. This represents the event-related input; each expert node performs local encoding on its corresponding information domain: , in, Indicates the first The initial hidden state of each node. This represents the local information input to this node; different nodes can use different network structures.

[0022] Preferably, in step four, after obtaining the initial state of each node, message passing is performed according to the ring topology; Each node passes messages to the next node: , in, Indicates the first In round propagation, nodes Passed to node Circular messages, Indicates the current hidden state of the node. This indicates the current prediction residual or risk characterization of this node. This indicates the confidence level or uncertainty information of the node; The next node receives a message from the previous node and updates its own state: , in, Indicates input up to the number The local information of the expert node, that is, the historical input sequence of the source-load joint prediction after being divided according to the information domain, and combined with the local information of the expert node. The sub-input corresponding to the information domain or prediction sub-task undertaken by each expert node; the local information, together with the node's current hidden state and the received cyclic messages, constitute the node state update network. The input from the last node is sent back to the first node, forming a closed loop: , , This closed propagation mechanism means that different expert nodes are no longer independent parallel branches, but rather form sequential interactions and state reflows within a unified ring structure.

[0023] Preferably, in step five, to prevent irrelevant information from repeatedly propagating in the ring network, this invention introduces a gating mechanism during the ring-shaped message passing process; for the first... Calculate the gating weight for each node's received messages: , in, Indicates message reception weight, This represents the node task query vector. This represents the Sigmoid function. This represents the weight matrix used to calculate the message reception gating weights, used for the... The current hidden state of each expert node Received loop messages and node task query vector The splicing result is linearly transformed, and the transformation result is compared with the bias term. After addition, the message reception gating weights are obtained by the Sigmoid function. , Indicates the first In the circular propagation of messages, the first... The expert node directs to the first The circular message transmitted by the expert node; the circular message is generated by the circular message generation network of the sending node based on the sending node's current hidden state, current prediction residual or risk characterization, and confidence or uncertainty information; when the first expert node transmits a circular message; When the first expert node is the first expert node, its preceding node is the last expert node in the ring network, and the corresponding received message is... ,in Total number of expert nodes; node status updated as follows: , in, Indicates the first The expert node at the _ In round-robin message propagation, the message reception gating weight is calculated for round-robin messages from the previous expert node; this weight is based on the... The candidate state is calculated from the current hidden state of each expert node, the received cyclic messages, and the node task query vector, and is used to control the candidate state after fusing the cyclic messages. With the current hidden state Hidden status after update The fusion ratio in It represents the candidate state after fusing the circular messages; when the information received by a node is irrelevant to its own task or has low confidence, the gating weight is reduced to suppress noise propagation; when the received information is highly correlated with the current prediction error, event risk or source-load coupling relationship, the gating weight is increased to enhance cross-node collaboration.

[0024] Preferably, in step six, to ensure that the multi-expert collaborative prediction results conform to the source-load physical relationship, a source-load consistency constraint is set; let the source-side prediction, load-side prediction, and net load prediction output by the model be as follows: , Then construct a balanced consistency loss: , This loss is used to ensure consistency between net load forecasts and source-side and load-side forecasts, avoiding inconsistencies in the overall result despite high accuracy of multiple forecast branches. Simultaneously, to prevent infinite state drift during circular propagation, a loop consistency constraint is constructed. After one complete round of circular propagation, the first node receives a return message from the last node. This return message should be consistent with the global source-load state and should not introduce irrelevant disturbances. Therefore, the loop consistency loss is defined as follows: , in, This represents the global state aggregation function; this constraint is used to ensure that multi-round circumferential propagation maintains the stability of the global predicted state while enhancing information interaction.

[0025] Preferably, the ring network in step seven can perform multiple rounds of propagation; let the maximum number of propagation rounds be... , No. Calculate the global benefit assessment value after round propagation: , in, This represents the return evaluation function, used to measure whether continuing to execute the next round of circular propagation can still significantly reduce prediction error or uncertainty; when the return is below a threshold... Stop spreading at that time: , when Or when the maximum number of propagation rounds is reached, the circular collaborative inference stops; this mechanism enables the model to complete predictions in fewer rounds on simple, stationary samples, and to perform more rounds of state interaction and residual correction on complex, abrupt, or strongly coupled samples. In step eight, after each expert node completes multiple rounds of collaboration, the prediction decoder outputs local prediction results: , Subsequently, the final source load prediction result is obtained through the fusion layer: , in, The output fusion network is represented by attention-weighted, confidence-weighted, gated fusion, or multilayer perceptron. The final output includes future source-side power output forecasts, load-side demand forecasts, net load forecasts, confidence scores for each forecast branch, key event risk scores, and interpretation results of the circumferential information flow. The training objectives in step nine include prediction error loss, source-load consistency loss, loop consistency loss, expert discrepancy loss, gating sparsity loss, and computational cost loss. , in, Used to constrain the accuracy of source-side, load-side, and net load forecasts; Used to constrain the source-load balance relationship; Used to constrain the state consistency of the circular propagation process; This is used to enhance the functional differences between different expert nodes and prevent all nodes from learning the same representation. Used to constrain the sparsity of gating weights and reduce the propagation of invalid information; Used to control the computational overhead caused by multi-round circular propagation; to Here, is the weighting factor; where, to The prediction error loss is determined by a fixed positive tradeoff coefficient through offline grid search. The coefficient is fixed at 1; the training set and validation set are divided according to the time sequence, and a finite set of candidate values ​​is set for each tradeoff coefficient. The Cartesian product of each set of candidate values ​​forms a set of candidate coefficient combinations. For each candidate coefficient combination, while maintaining the network structure and other training configurations, the ring-coupled multi-expert neural network is trained using the training set. A unified prediction error evaluation index is calculated based on the source-side output, load-side demand, and net load prediction results of the trained network on the validation set. The candidate coefficient combination with the smallest evaluation index is selected as the determined tradeoff coefficient combination. The tradeoff coefficient remains fixed in each candidate training and the selected model training process, and does not change with the training batch, the number of training iterations, or the number of rounds of circumferential propagation, and is not used as a network parameter updated through backpropagation.

[0026] The present invention also provides a source-load joint prediction system based on a ring-coupled multi-expert neural network that employs the above-mentioned prediction method. The system can be deployed in any of the following scenarios to perform corresponding source-load prediction tasks: deployed in a new energy power station for wind and solar power prediction; deployed in an integrated energy system for multi-energy flow prediction of cooling, heating and power; deployed in a virtual power plant for adjustable resource aggregation prediction; deployed in a park microgrid for local supply and demand balance prediction; and deployed in a source-grid-load-storage coordinated scheduling system for multi-timescale scheduling decisions.

[0027] By adopting the above structure, this invention has the following beneficial effects: Compared with existing single-backbone deep learning prediction models, this invention decomposes the source-load prediction task into multiple expert nodes with specialized responsibilities, and achieves multi-round interaction through a circular topology. Therefore, it can better express the complex coupling relationship between the source side, load side, meteorology, time, operating status, and event risk. Traditional methods usually perform one-time fusion at the input or output end, while this invention achieves stepwise collaborative reasoning through circular message passing, enabling source-side prediction errors, load-side peak-valley variations, meteorological abrupt changes, and net load risk to be cyclically transmitted and corrected among multiple nodes.

[0028] Compared to existing multi-task learning methods, this invention not only shares features but also explicitly models the directional information flow and residual backflow paths between tasks. Source-side experts, load-side experts, meteorological experts, and event risk experts can perform multiple rounds of state updates, thus avoiding the problem of independent and insufficiently coordinated outputs of task branches in traditional multi-task learning.

[0029] Compared to existing expert hybrid methods, the expert nodes in this invention are not simply parallel modules selected by a gating network, but rather form a closed collaborative structure according to a ring topology. Each expert retains its own specialized modeling capabilities while participating in the state updates of other experts through circumferential message passing, thus enabling stronger cross-domain coupled modeling capabilities and a clearer information flow interpretation path.

[0030] Compared to existing graph neural network methods, this invention does not rely on a fixed power grid topology or manually constructed fully connected graphs. Instead, it constructs a ring-shaped collaborative structure based on the inherent information domain relationships of the source-load prediction task. This structure is not only more suitable for expressing the task coupling relationships in source-load prediction than a fixed graph, but also offers greater computational controllability and interpretability than a fully connected attention structure.

[0031] Compared to existing net load forecasting methods, this invention explicitly introduces source-load balance consistency constraints to maintain coordination among source-side output, load-side demand, and net load forecasts. This design reduces inconsistencies between multiple output forecasts and improves the model's usability in scheduling decisions.

[0032] Compared to conventional deep learning models, this invention can also output the circular propagation path, expert gating weights, node confidence, and event risk score, thereby enhancing the interpretability of the prediction results. Schedulers can not only obtain future source-load predictions, but also understand whether the current prediction is mainly affected by meteorological disturbances, load cycles, source-side fluctuations, or event risks.

[0033] In summary, this invention proposes a source-load joint prediction method based on a ring-coupled multi-expert neural network. This method organizes multiple neural network sub-models with different prediction responsibilities into a closed ring topology. Each sub-network node performs local modeling for specific information domains such as source-side output, load-side demand, meteorological disturbances, time periods, operating status, and event risk or uncertainty modeling. Nodes exchange hidden states, residual information, and confidence information sequentially through a ring-shaped message passing mechanism, forming a globally consistent source-load prediction result after multiple rounds of ring propagation. Thus, the prediction process is no longer a single feature fusion but is transformed into a closed-loop collaborative reasoning process of "multi-expert local modeling—ring-shaped information passing—residual backflow correction—source-load consistency constraints—adaptive output fusion." Furthermore, the method and corresponding system in this invention are well-suited for scenarios such as wind power prediction, photovoltaic power prediction, electricity load prediction, net load prediction, multi-load prediction of integrated energy systems, source-load prediction of industrial microgrids, joint prediction of virtual power plant output and load, and coordinated scheduling prediction of source, grid, load, and storage. Attached Figure Description

[0034] Figure 1 This is a ring topology diagram of the source-side, load-side, meteorological, time, operational status, and event risk expert nodes involved in this invention.

[0035] Figure 2 This is a schematic diagram of the internal structure of a single expert node in an embodiment of the present invention.

[0036] Figure 3 This is a flowchart of multi-round circular collaborative reasoning and adaptive stopping in an embodiment of the present invention.

[0037] Figure 4 This is a flowchart of the gating message passing and state update mechanism of node i in an embodiment of the present invention.

[0038] Figure 5 This is a schematic diagram of source-load balance consistency constraints and loop dual consistency constraints in an embodiment of the present invention. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings.

[0040] This invention provides a source-load joint prediction method and system based on a ring-coupled multi-expert neural network. The core idea is to use multiple neural network sub-models with different modeling responsibilities as prediction nodes in a ring topology. Each node is responsible for a specific information domain or prediction sub-task, and multiple rounds of state exchange and residual correction are performed through a ring-shaped message passing mechanism, ultimately outputting the joint prediction results of source-side output, load-side demand, and net load.

[0041] The specific implementation includes the following steps: Step 1: Obtain and construct the historical input sequence for joint source-load forecasting; obtain source-side output data, load-side demand data, meteorological data, time calendar data, operational status data, and event-related data for the target area or target station, align them according to a unified sampling interval, and construct the historical input sequence; divide the corresponding local sub-inputs according to the information domain to which each type of data belongs, and determine the source-side output, load-side demand, and net load forecasting tasks within the future forecast interval; Step Two: Construct a ring-coupled multi-expert neural network node; Based on the information domain modeling responsibilities of source-side output, load-side demand, meteorological disturbances, time rhythm, operating status, and event risk, set up expert neural network nodes, connect them in a preset order, so that the message output of each node is connected to the message input of the next node, and the message output of the last node is connected to the message input of the first node, forming a closed ring topology; each expert neural network node is a neural network sub-model including a local encoder, a node state update network, a ring-shaped message generation network, and a node predictive decoder, represented as: , in, The total number of expert nodes. Used to convert the corresponding local sub-input into a hidden state. Used to generate candidate states based on the current hidden state, received messages, and local information. Used to generate a cyclic message to be passed to the next node. Used to convert the hidden state of a node into a local prediction result; Step 3: Each expert node performs local encoding; each local sub-input is input into the local encoder of the corresponding expert node to extract the temporal features of that information domain, forming the initial hidden state of each expert node: , in, For the first Each expert node has its own local sub-input; each expert node performs local representation learning on its own information domain and uses the initial hidden state as the initial state for circular collaborative reasoning. Step 4: Construct a circular message passing mechanism; the circular message generation network of the sending node combines and transforms the node's current hidden state, prediction residual or risk characterization, and confidence or uncertainty information to generate a circular message, which is then passed to the next node along the closed ring topology; the receiving node inputs its current hidden state, the received circular message, and its own local information into the node state update network to generate a candidate state after fusing the circular message; the last node sends its circular message back to the first node to achieve information feedback; Step 5: Construct a cyclic gating and conflict suppression mechanism; concatenate the current hidden state of the receiving node, the received cyclic message, and the task query vector representing the node's modeling responsibility, and obtain the message receiving gating weight through linear transformation and the Sigmoid function; according to the gating weight, perform element-wise weighted fusion of the candidate state and the original hidden state of the receiving node to obtain the updated hidden state; by controlling the candidate state update ratio caused by the cyclic message, suppress the interference of information unrelated to the node's task or with low credibility on the current predicted state; Step Six: Construct source-load consistency and loop consistency constraints; construct the source-load balance consistency loss based on source-side forecasts, load-side forecasts, and net load forecasts: , in, , and These are the predicted sequences for source-side output, load-side demand, and net load, respectively. Based on the hidden states of expert nodes in adjacent rounds, a loop consistency loss is constructed using a global state aggregation function. , in, This is a global state aggregation function. For the propagation rounds; the above two losses are used to constrain the balance deviation between the source load prediction results and the changes in the global state before and after circumferential propagation, respectively; Step 7: Perform multi-round circular collaborative reasoning and adaptive stopping; repeatedly execute circular message generation, message passing, candidate state generation and gating state update, and calculate the benefit evaluation value of continuing propagation based on the current hidden state of each expert node after each round of propagation; when the benefit evaluation value is not higher than the preset threshold, or when the number of propagation rounds reaches the preset maximum number of propagation rounds, stop the circular collaborative reasoning and retain the final hidden state of each expert node. Step 8: Output source-side, load-side, and net load prediction results; input the final hidden state of each expert node into the corresponding node prediction decoder to obtain local prediction results, and fuse the collaboratively formed prediction information through the output fusion network to obtain the source-side output prediction sequence, load-side demand prediction sequence, and net load prediction sequence within the future prediction interval. Step Nine: Construct a Joint Training Objective; During the training phase, use historical input sequences and their corresponding prediction intervals to construct training samples and establish a joint training objective: , in, To predict error loss, For source load balance consistency loss, For loop consistency loss, For expert difference loss, For gating sparsity loss, To calculate the cost loss, to The weighting coefficients for each loss term are used; the trainable parameters involved in local encoding, state update, message generation, node prediction decoding, gating, benefit evaluation, and output fusion are jointly optimized using the joint training objective; during the prediction phase, the trained network is invoked to perform the corresponding encoding, collaborative inference, and output operations.

[0042] Combining the above methods and steps, and further combining the appendix Figure 1 To the attached Figure 5 The specific execution of each step is further explained in detail in the form of an embodiment as follows: Step 1: Obtain and construct the historical input sequence for joint source-load prediction.

[0043] First, acquire historical multivariate time series data for the target area or target power station. Input data includes, but is not limited to, source-side power output data, load-side demand data, meteorological data, time calendar data, power grid operation status data, and external event data.

[0044] Source-side power output data may include wind power, photovoltaic power, distributed renewable energy output, and energy storage charging and discharging power; load-side demand data may include total load, industrial load, commercial load, residential load, charging pile load, and park load; meteorological data may include wind speed, wind direction, irradiance, temperature, humidity, air pressure, precipitation, and cloud cover; time calendar data may include hour, day type, weekday, holiday, season, and month; and operational status data may include line power flow, node voltage, energy storage status, demand response status, and day-ahead plans.

[0045] The above data is time-aligned according to a uniform sampling interval to construct a historical input sequence: , in, Indicates the length of the history window. Indicates time Multivariate inputs. Prediction targets include future... Step-by-step power output sequence, load-side demand sequence, and net load sequence: , in, Indicates the power output from the source side. Indicates load-side demand. This represents the net load. Generally, the net load can be determined by the difference between load-side demand and source-side renewable energy output, i.e.: , In practical applications, the definition of net load can also be expanded to include factors such as energy storage, purchased electricity, and controllable load.

[0046] Step 2: Construct a ring-coupled multi-expert neural network node.

[0047] This embodiment breaks down the source load prediction task into several complementary information domain modeling tasks, and constructs an expert neural network node for each type of task. Multiple expert nodes are connected in a ring topology to form a closed cooperative network.

[0048] In a preferred embodiment, the ring network includes the following six types of expert nodes: 1. Source-side power output expert node, used to model the historical changes, ramp-up events, and meteorological driving relationships of power output from new energy sources such as wind power and photovoltaics; 2. Load-side demand expert node, used to model the periodicity, peak-valley characteristics, holiday effects, and changes in user behavior of load demand; 3. Meteorological disturbance expert node, used to model the combined effects of meteorological variables on source-side output and load-side demand; 4. Time rhythm expert nodes, used to model daily, weekly, seasonal, and holiday patterns; 5. Operational status expert node, used to model the impact of source-grid-load-storage operation status, energy storage charging and discharging, demand response, and scheduling plans on future source-load changes; 6. Event Risk Expert Node: Used to identify high-risk prediction segments such as source-side ramp-up, sudden load increase, meteorological changes, and sharp changes in net load.

[0049] The number of expert nodes mentioned above is not a limitation. In different application scenarios, it can be set to four, five, six or more nodes; nodes can also be deleted or added according to the actual availability of data.

[0050] Assume the ring network has a total of The expert node, the first Each expert node is represented as: , in, Indicates a local encoder. This indicates that the node status is updated in the network. This indicates a circular message generation network. This represents the node prediction decoder.

[0051] Step 3: Each expert node performs local coding.

[0052] Divide the input sequence into different sub-inputs according to the information domain: , in, Indicates the source-side input. Indicates load-side input. Indicates meteorological input. This indicates the input of a time calendar. Indicates the running status input. This indicates input related to the event.

[0053] Each expert node performs local encoding on its corresponding information domain: , in, Indicates the first The initial hidden state of each node. This represents the local information input to this node. Different nodes can use different network structures. For example, source-side output experts can use temporal convolutional networks or Transformer structures, load-side demand experts can use periodic attention networks, meteorological disturbance experts can use multivariate cross-attention networks, and event risk experts can use rate-of-change sensing networks or uncertainty sensing networks.

[0054] Through this step, different information domains first complete local representation learning in their respective suitable model structures, avoiding information aliasing caused by simply splicing all variables at the input end.

[0055] Step 4: Build a circular message passing mechanism.

[0056] After obtaining the initial state of each node, this invention performs message passing according to a ring topology. Each node passes a message to the next node: , in, Indicates the first In round propagation, nodes Passed to node Circular messages, Indicates the current hidden state of the node. This indicates the current prediction residual or risk characterization of this node. This indicates the confidence level or uncertainty information of the node.

[0057] The next node receives a message from the previous node and updates its own state: , in, Indicates input up to the number The local information of the expert node, that is, the historical input sequence of the source-load joint prediction after being divided according to the information domain, and combined with the local information of the expert node. The sub-input corresponding to the information domain or prediction sub-task undertaken by each expert node; the local information, together with the node's current hidden state and the received cyclic messages, constitute the node state update network. The input from the last node is sent back to the first node, forming a closed loop: , , This closed-loop propagation mechanism transforms the independent parallel branches between different expert nodes into sequential interactions and state feedback within a unified ring structure. For example, a weather disturbance node can transmit wind speed change information to the source-side output node, which in turn transmits the predicted residual information to the net load-related node. The load node then transmits peak-valley variation information to the event risk node, which in turn feeds back high-risk segment information to the source or load-side node, thus forming multiple rounds of collaborative correction.

[0058] Step 5: Construct a cyclic gating and conflict suppression mechanism.

[0059] To prevent irrelevant information from repeatedly propagating in a ring network, this invention introduces a gating mechanism during the ring-shaped message passing process. For the first... Calculate the gating weight for each node's received messages: , in, Indicates message reception weight, This represents the node task query vector. This represents the Sigmoid function. This represents the weight matrix used to calculate the message reception gating weights, used for the... The current hidden state of each expert node Received loop messages and node task query vector The splicing result is linearly transformed, and the transformation result is compared with the bias term. After addition, the message reception gating weights are obtained by the Sigmoid function. , Indicates the first In the circular propagation of messages, the first... The expert node directs to the first The circular message transmitted by the expert node; the circular message is generated by the circular message generation network of the sending node based on the sending node's current hidden state, current prediction residual or risk characterization, and confidence or uncertainty information; when the first expert node transmits a circular message; When the first expert node is the first expert node, its preceding node is the last expert node in the ring network, and the corresponding received message is... ,in Total number of expert nodes; node status updated as follows: , in, Indicates the first The expert node at the _ In round-robin message propagation, the message reception gating weight is calculated for round-robin messages from the previous expert node; this weight is based on the... The candidate state is calculated from the current hidden state of each expert node, the received cyclic messages, and the node task query vector, and is used to control the candidate state after fusing the cyclic messages. With the current hidden state Hidden status after update The fusion ratio in This indicates the candidate state after merging the circular messages.

[0060] When the information received by a node is irrelevant to its own task or has low confidence, the gating weight is reduced to suppress noise propagation; when the received information is highly correlated with the current prediction error, event risk, or source-load coupling relationship, the gating weight is increased to enhance cross-node collaboration.

[0061] Step 6: Construct source-load consistency and loop consistency constraints.

[0062] To ensure that the multi-expert collaborative prediction results conform to the physical relationship between the source and load, this invention sets source-load consistency constraints. Let the source-side prediction, load-side prediction, and net load prediction output by the model be as follows: , Then construct a balanced consistency loss: , This loss is used to ensure consistency between net load forecasts and source-side and load-side forecasts, avoiding situations where multiple forecast branches have high individual accuracy but the overall results are inconsistent.

[0063] Meanwhile, to avoid infinite state drift during circular propagation, this invention constructs a loop consistency constraint, as shown in the appendix. Figure 5 As shown, source-load consistency constraints and loop consistency constraints constitute a dual consistency constraint. After one complete round of loop propagation, the first node receives a return message from the last node. This return message should maintain consistency with the global source-load state and should not introduce irrelevant disturbances. Therefore, the loop consistency loss is defined as follows: , in, This represents the global state aggregation function. This constraint is used to ensure that multi-round circumferential propagation maintains the stability of the global predicted state while enhancing information interaction.

[0064] Step 7: Perform multi-round circular collaborative reasoning and adaptive stopping.

[0065] A ring network can perform multiple rounds of propagation. Let the maximum number of propagation rounds be... , No. Calculate the global benefit assessment value after round propagation: , in, This represents the return evaluation function, used to measure whether continuing to execute the next round of circular propagation can still significantly reduce prediction error or uncertainty. When the return is below a threshold... Stop spreading at that time: , when The circular collaborative inference stops when the maximum number of propagation rounds is reached. This mechanism enables the model to complete predictions in fewer rounds on simple, stationary samples, and to perform more rounds of state interaction and residual correction on complex, abrupt, or strongly coupled samples.

[0066] Step 8: Output the source-side, load-side, and net load prediction results.

[0067] After each expert node completes multiple rounds of collaboration, the prediction decoder outputs local prediction results: , Subsequently, the final source load prediction result is obtained through the fusion layer: , in, The output fusion network can be implemented using attention-weighted, confidence-weighted, gated fusion, or multilayer perceptron. The final output includes future source-side power output forecasts, load-side demand forecasts, net load forecasts, confidence scores for each forecast branch, key event risk scores, and interpretation results of the cyclic information flow.

[0068] Step 9: Construct joint training objectives.

[0069] The training objectives of this invention include prediction error loss, source-load consistency loss, loop consistency loss, expert discrepancy loss, gating sparsity loss, and computational cost loss. , in, Used to constrain the accuracy of source-side, load-side, and net load forecasts; Used to constrain the source-load balance relationship; Used to constrain the state consistency of the circular propagation process; This is used to enhance the functional differences between different expert nodes and prevent all nodes from learning the same representation. Used to constrain the sparsity of gating weights and reduce the propagation of invalid information; Used to control the computational overhead caused by multi-round circular propagation; to Here, is the weighting factor; where, to The prediction error loss is determined by a fixed positive tradeoff coefficient through offline grid search. The coefficient is fixed at 1; the training set and validation set are divided according to the time sequence, and a finite set of candidate values ​​is set for each tradeoff coefficient. The Cartesian product of each set of candidate values ​​forms a set of candidate coefficient combinations. For each candidate coefficient combination, while maintaining the network structure and other training configurations, the ring-coupled multi-expert neural network is trained using the training set. A unified prediction error evaluation index is calculated based on the source-side output, load-side demand, and net load prediction results of the trained network on the validation set. The candidate coefficient combination with the smallest evaluation index is selected as the determined tradeoff coefficient combination. The tradeoff coefficient remains fixed in each candidate training and the selected model training process, and does not change with the training batch, the number of training iterations, or the number of rounds of circumferential propagation, and is not used as a network parameter updated through backpropagation.

[0070] After implementation and verification, the key improvements of this invention include the following.

[0071] First, an improved ring-coupled multi-expert neural network structure for joint source-load prediction is proposed. This structure connects multiple neural network nodes with different responsibilities in a closed ring topology, instead of using ordinary parallel integration, simple splicing fusion, or unidirectional hierarchical fusion.

[0072] Second, the expert node partitioning method has been improved for the source side, load side, meteorological side, time side, operational status side, and event risk side. Each node has an independent encoder, state update network, message generation network, and prediction decoder, enabling specialized modeling for different information domains.

[0073] Third, the circular message passing mechanism has been improved. This mechanism enables each expert node to pass its hidden state, prediction residuals, risk characterization, and confidence information to neighboring nodes, and achieves cross-task information exchange in multiple rounds of propagation.

[0074] Fourth, the circumferential gating and conflict suppression mechanism has been improved. This mechanism calculates the gating weight based on the current node state, received messages, and task query vectors, thereby suppressing the propagation of irrelevant information and enhancing the effective coupling of information transmission.

[0075] Fifth, the source-load balance consistency constraint has been improved. This constraint explicitly guarantees that the source-side forecast, load-side forecast, and net load forecast satisfy a physical consistency relationship, avoiding contradictions between multiple output forecast results.

[0076] Sixth, the loop consistency constraint has been improved. This constraint is used to control the global state drift after loop propagation, enabling multi-round loop interactions to maintain prediction stability while improving collaborative modeling capabilities.

[0077] Seventh, an improved adaptive stopping mechanism based on propagation benefit evaluation was implemented. This mechanism can determine the number of circular propagation rounds according to the sample complexity, reducing inference overhead for simple samples and enabling more comprehensive multi-node collaborative correction for complex samples.

[0078] Eighth, adaptable to various application scenarios, based on the above-mentioned ring-coupled multi-expert network to realize the synchronous prediction of source-side output, load-side demand and net load, and the application of this system in new energy power stations, integrated energy systems, virtual power plants, park microgrids and source-grid-load-storage coordinated scheduling.

[0079] In new power systems with a high proportion of renewable energy integration, the output from sources such as wind and solar power exhibits significant randomness, volatility, and weather-driven characteristics, while load-side demand is influenced by multiple factors including temperature, holidays, user behavior, industrial structure, and demand response. Source-side output and load-side demand are not two independent predictable objects, but rather jointly determine the system's equilibrium state through dispatching plans, reserve configuration, net load changes, coordinated operation of power generation, grid, load, and storage, and market clearing. Therefore, predicting renewable energy output or load demand separately is insufficient to fully characterize the dynamic coupling relationship between sources and load.

[0080] This invention proposes a source-load joint prediction method based on a ring-coupled multi-expert neural network. It organizes multiple neural network sub-models with different prediction responsibilities into a closed ring topology. Each sub-network node performs local modeling for specific information domains such as source-side output, load-side demand, meteorological disturbances, time periods, operating status, and event risk or uncertainty modeling. Nodes exchange hidden states, residual information, and confidence information sequentially through a ring-shaped message passing mechanism, forming a globally consistent source-load prediction result after multiple rounds of ring propagation. Thus, the prediction process is no longer a single feature fusion but is transformed into a closed-loop collaborative reasoning process of "multi-expert local modeling—ring-shaped information passing—residual backflow correction—source-load consistency constraints—adaptive output fusion." Implementation verification has shown that this method effectively solves the aforementioned technical problems.

[0081] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual technical solution is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar solutions and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A source-load joint prediction method based on a ring-coupled multi-expert neural network, characterized in that: Multiple neural network sub-models with different modeling responsibilities are used as prediction nodes in a ring topology. Each node is responsible for a specific information domain or prediction sub-task. Multiple rounds of state exchange and residual correction are performed through a ring-shaped message passing mechanism, and finally the joint prediction results of source-side output, load-side demand and net load are output. The specific prediction method includes the following steps: Step 1: Obtain and construct the historical input sequence for joint source-load prediction; Step 2: Construct ring-coupled multi-expert neural network nodes; Step 3: Each expert node performs local coding; Step 4: Establish a circular message passing mechanism; Step 5: Construct a cyclic gating and conflict suppression mechanism; Step Six: Construct source-load consistency and loop consistency constraints; Step 7: Perform multi-round circular collaborative reasoning and adaptive stopping; Step 8: Output the source-side, load-side, and net load prediction results; Step 9: Construct joint training objectives.

2. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: Step one first acquires historical multivariate time series data for the target area or target power station. Input data includes, but is not limited to, source-side output data, load-side demand data, meteorological data, time calendar data, power grid operation status data, and external event data. Source-side output data includes wind power, photovoltaic power, distributed renewable energy output, and energy storage charging / discharging power. Load-side demand data includes total load, industrial load, commercial load, residential load, charging pile load, and park load. Meteorological data includes wind speed, wind direction, irradiance, temperature, humidity, air pressure, precipitation, and cloud cover. Time calendar data may include hour, day type, weekday, holiday, season, and month. Operation status data may include line power flow, node voltage, energy storage status, demand response status, and day-ahead plan. The above data are then time-aligned according to a unified sampling interval to construct a historical input sequence. ， in, Indicates the length of the history window. Indicates time Multivariate inputs; prediction targets include future... Step-by-step power output sequence, load-side demand sequence, and net load sequence: ， in, Indicates the power output from the source side. Indicates load-side demand. This represents the net load; the net load is determined by the difference between load-side demand and source-side renewable energy output, i.e.: .

3. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step two, the source load prediction task is divided into several complementary information domain modeling tasks, and an expert neural network node is constructed for each type of task; multiple expert nodes are connected in a ring topology to form a closed cooperative network. The ring network includes: source-side output expert nodes, used to model historical changes, ramp-up events, and meteorological driving relationships of wind and solar power output; load-side demand expert nodes, used to model the periodicity, peak-valley characteristics, holiday effects, and user behavior changes of load demand; meteorological disturbance expert nodes, used to model the combined impact of meteorological variables on source-side output and load-side demand; time rhythm expert nodes, used to model daily, weekly, seasonal, and holiday patterns; operation status expert nodes, used to model the impact of source-grid-load-storage operation status, energy storage charging and discharging, demand response, and scheduling plans on future source-load changes; and event risk expert nodes, used to identify high-risk prediction segments such as source-side ramp-up, load surges, meteorological abrupt changes, and sharp changes in net load. In different application scenarios, four, five, six, or more nodes can be configured; or nodes can be added or removed based on actual data availability. Assume the ring network has a total of The expert node, the first Each expert node is represented as: ， in, Indicates a local encoder. This indicates that the node status is updated in the network. This indicates a circular message generation network. This represents the node prediction decoder.

4. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step three, the input sequence is divided into different sub-inputs according to the information domain: ， in, Indicates the source-side input. Indicates load-side input. Indicates meteorological input. This indicates the input of a time calendar. Indicates the running status input. This represents the event-related input; each expert node performs local encoding on its corresponding information domain: ， in, Indicates the first The initial hidden state of each node. This represents the local information input to this node; different nodes can use different network structures.

5. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step four, after obtaining the initial state of each node, message passing is performed according to the ring topology; Each node passes a message to the next node: ， in, Indicates the first In round propagation, nodes Passed to node Circular messages, Indicates the current hidden state of the node. This indicates the current prediction residual or risk characterization of this node. This indicates the confidence level or uncertainty information of the node; The next node receives a message from the previous node and updates its own state: ， in, Indicates input up to the number The local information of the first expert node, that is, the source-load joint prediction input sequence after being divided according to the information domain, and its relationship with the first expert node. The sub-input corresponding to the information domain or prediction sub-task undertaken by each expert node; the local information, together with the node's current hidden state and the received cyclic messages, constitute the node state update network. The input from the last node is sent back to the first node, forming a closed loop: ， ， This closed propagation mechanism ensures that different expert nodes are no longer independent parallel branches, but rather form sequential interactions and state reflows within a unified ring structure.

6. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step five, to prevent irrelevant information from repeatedly propagating in the ring network, a gating mechanism is introduced during the ring-shaped message passing process; for the first... Calculate the gating weight for each node's received messages: ， in, Indicates message reception weight, This represents the node task query vector. This represents the Sigmoid function. This represents the weight matrix used to calculate the message reception gating weights, used for the... The current hidden state of each expert node Received loop messages and node task query vector The splicing result is linearly transformed, and the transformation result is compared with the bias term. After addition, the message reception gating weights are obtained by the Sigmoid function. , Indicates the first In the circular propagation of messages, the first... The expert node directs to the first The circular message transmitted by the expert nodes; the circular message is generated by the circular message generation network of the sending node based on the sending node's current hidden state, current prediction residual or risk characterization, and confidence or uncertainty information; when the first expert node transmits a circular message; When the first expert node is the first expert node, its preceding node is the last expert node in the ring network, and the corresponding received message is... ,in Total number of expert nodes; node status updated as follows: ， in, Indicates the first The expert node at the _ In round-robin message propagation, the message reception gating weight is calculated for round-robin messages from the previous expert node; this weight is based on the... The candidate state is calculated from the current hidden state of each expert node, the received cyclic messages, and the node task query vector, and is used to control the candidate state after fusing the cyclic messages. With the current hidden state Hidden status after update The fusion ratio in It represents the candidate state after fusing the circular messages; when the information received by a node is irrelevant to its own task or has low confidence, the gating weight is reduced to suppress noise propagation; when the received information is highly correlated with the current prediction error, event risk or source-load coupling relationship, the gating weight is increased to enhance cross-node collaboration.

7. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step six, to ensure that the multi-expert collaborative prediction results conform to the source-load physical relationship, a source-load consistency constraint is set; let the source-side prediction, load-side prediction, and net load prediction output by the model be as follows: ， Then construct a balanced consistency loss: ， This loss is used to ensure consistency between net load forecasts and source-side and load-side forecasts, avoiding inconsistencies in the overall results despite high accuracy of multiple forecast branches. At the same time, to prevent infinite state drift during the circular propagation process, a loop consistency constraint is constructed. After a complete round of circular propagation, the first node receives a return message from the last node. This return message should be consistent with the global source load state and should not introduce irrelevant disturbances. Therefore, the loop consistency loss is defined as follows: ， in, This represents the global state aggregation function; this constraint is used to ensure that multi-round circumferential propagation maintains the stability of the global predicted state while enhancing information interaction.

8. The source-load joint prediction method based on a ring-coupled multi-expert neural network according to claim 1, characterized in that: In step seven, the ring network can perform multiple rounds of propagation; let the maximum number of propagation rounds be... , No. Calculate the global benefit assessment value after round propagation: ， in, This represents the return evaluation function, used to measure whether continuing to execute the next round of circular propagation can still significantly reduce prediction error or uncertainty; when the return is below a threshold... Stop spreading at that time: ， when Or when the maximum number of propagation rounds is reached, the circular collaborative inference stops; this mechanism enables the model to complete predictions in fewer rounds on simple, stationary samples, and to perform more rounds of state interaction and residual correction on complex, abrupt, or strongly coupled samples; In step eight, after each expert node completes multiple rounds of collaboration, the prediction decoder outputs local prediction results: ， Subsequently, the final source load prediction result is obtained through the fusion layer: ， in, The output fusion network is represented by attention-weighted, confidence-weighted, gated fusion, or multilayer perceptron. The final output includes future source-side power output forecasts, load-side demand forecasts, net load forecasts, confidence scores for each forecast branch, key event risk scores, and interpretation results of the circumferential information flow. The training objectives in step nine include prediction error loss, source-load consistency loss, loop consistency loss, expert discrepancy loss, gating sparsity loss, and computational cost loss. ， in, Used to constrain the accuracy of source-side, load-side, and net load forecasts; Used to constrain the source-load balance relationship; Used to constrain the state consistency of the circular propagation process; This is used to enhance the functional differences between different expert nodes and prevent all nodes from learning the same representation. Used to constrain the sparsity of gating weights and reduce the propagation of invalid information; Used to control the computational overhead caused by multi-round circular propagation; to Here, is the weighting factor; where, to For the fixed positive tradeoff coefficients determined through offline grid search, the prediction error loss is... The coefficient is fixed at 1; the training set and validation set are divided according to the time sequence, and a finite set of candidate values ​​is set for each tradeoff coefficient. The Cartesian product of each set of candidate values ​​forms a set of candidate coefficient combinations. For each candidate coefficient combination, while maintaining the network structure and other training configurations, the ring-coupled multi-expert neural network is trained using the training set. A unified prediction error evaluation index is calculated based on the source-side output, load-side demand, and net load prediction results of the trained network on the validation set. The candidate coefficient combination with the smallest evaluation index is selected as the determined tradeoff coefficient combination. The tradeoff coefficient remains fixed in each candidate training and the selected model training process, and does not change with the training batch, the number of training iterations, or the number of rounds of circumferential propagation, and is not used as a network parameter updated through backpropagation.

9. A source-load joint prediction system based on a ring-coupled multi-expert neural network, employing the method described in any one of claims 1-8, characterized in that: The system can be deployed in any of the following scenarios to perform the corresponding source-load prediction tasks: deployed at new energy power plants for wind and solar power prediction; Deployed in integrated energy systems for predicting multiple energy flows, including cooling, heating, and electricity; Deployed in virtual power plants for adjustable resource aggregation and prediction; Deployed in the park's microgrid for local supply and demand balance forecasting; deployed in the source-grid-load-storage coordinated dispatch system for multi-timescale dispatch decisions.

Images