New energy output prediction method and system based on structural plasticity
By simulating the structural plasticity mechanism of the biological nervous system, the synaptic death and growth of the new energy output prediction model are dynamically adjusted, which solves the problem of high computation and storage costs of the new energy output prediction model and realizes efficient and robust new energy output prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-04-03
AI Technical Summary
Existing new energy power output prediction models have high computation and storage costs in large-scale grid-connected scenarios, and traditional pruning methods cannot adapt to the random fluctuations in new energy power output and the dynamic changes in data distribution, resulting in decreased model performance and increased training costs.
By simulating the structural plasticity mechanism of biological nervous systems, a dynamic equilibrium temporal network model is constructed by dynamically adjusting the death and growth of synapses during network training, thereby optimizing the network structure and achieving model compression.
It achieves efficient prediction of new energy output, adapts to dynamic changes in data distribution, improves model robustness and prediction accuracy, and reduces computation and storage costs.
Smart Images
Figure CN121787634A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning technology, and in particular relates to a new energy output prediction method and system based on structural plasticity. Background Technology
[0002] With the rapid development of renewable energy, the high uncertainty brought about by large-scale renewable energy grid connection has become a key challenge restricting the stable operation of new power systems. Accurate power output forecasting is a crucial aspect of grid dispatching and is of great significance for improving renewable energy absorption capacity and reducing system operating costs. In recent years, deep learning technology has achieved remarkable results in the field of time series forecasting, especially recurrent neural networks such as LSTM and GRU and their variants, which have performed exceptionally well in renewable energy power output forecasting. However, with large-scale renewable energy grid connection, renewable energy power output forecasting models exhibit characteristics of large model size, high computational load, and high parameter count, thereby increasing the computational and storage costs of the models and making them difficult to deploy on resource-constrained edge devices. Therefore, how to construct lightweight and efficient renewable energy power output forecasting models is a key focus at present.
[0003] Model pruning, as an effective model compression method, reduces model complexity and computational overhead by removing redundant parameters from neural networks, while preserving model performance as much as possible. Traditional pruning methods are mainly divided into two categories: structured pruning and unstructured pruning. Structured pruning typically removes entire channels, filters, or layers, while unstructured pruning can remove parameters at arbitrary locations, achieving higher compression ratios.
[0004] However, traditional pruning methods have the following limitations in the scenario of predicting renewable energy output: First, most existing mainstream pruning methods are static pruning strategies that are either one-time or phased, and the network structure after pruning remains unchanged, which cannot adapt to the random fluctuations of renewable energy output and the dynamic changes in data distribution; Second, traditional pruning methods mainly make pruning decisions based on single indicators such as weight size and gradient information, ignoring the use of model parameters in long-term prediction tasks; Third, the pruning process is often accompanied by a decline in model performance, requiring additional fine-tuning to improve performance, which also increases training costs.
[0005] In recent years, inspired by the structural plasticity mechanism of biological neural systems, researchers have begun to explore dynamic pruning methods. Synapses in biological neural systems continuously adjust their connection strength based on usage; synapses that remain inactive for extended periods are weakened or even disappear, while highly active synapses are strengthened. This structural plasticity mechanism allows networks to adaptively adjust their structure, improving learning efficiency, and is particularly suitable for new energy output prediction tasks with dynamically changing data distributions.
[0006] Therefore, by simulating the structural plasticity learning mechanism of biological nervous systems, we study the synaptic growth and death mechanism during network training to achieve network structure optimization and lightweight, efficient operation of the new energy output prediction model. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a new energy output prediction method and system based on structural plasticity. Targeting the task of new energy output prediction, this invention proposes a method that simulates the structural plasticity of biological nervous systems, mimics the death and growth of synapses during network training, and implements this in a temporal network model to achieve network structure optimization and model compression.
[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0009] In a first aspect, the present invention provides a new energy output prediction method based on structural plasticity, the method comprising:
[0010] Obtain historical, timestamped renewable energy output data and manually annotate it;
[0011] The power data is preprocessed and then divided into training set, validation set and test set according to the proportions.
[0012] A dynamic equilibrium time series network model is constructed, and it is trained, validated, and tested using training, validation, and test sets.
[0013] A well-trained, validated, and tested dynamic equilibrium time series network model is used for predicting the output of new energy sources.
[0014] The dynamically balanced temporal network model refers to the method by which neurons in existing temporal network models maintain network dynamic stability through the dynamic death and growth characteristics of synapses; specifically:
[0015] Activation characteristic values are calculated based on synaptic gradients;
[0016] Calculate the survival time of each synapse, and construct a survival time control mechanism based on the synapse survival time;
[0017] The probability of synapse extinction is calculated based on activation characteristic values and survival time control mechanisms;
[0018] Synaptic extinction and regeneration are dynamically adjusted based on extinction probability and activation characteristic values.
[0019] Preferably, the process of dynamically adjusting synaptic extinction and regeneration based on the extinction probability and activation characteristic value includes:
[0020] Each synapse dies with a probability of demise, and the probability of demise is positively correlated with the synapse demise.
[0021] After the synapse dies, the required number of regenerated synapses is obtained based on the target sparsity. ; The required regenerated synapses regenerate new synapses through a random growth strategy. The growth ratio is random; the remaining number of synapses with the highest activation value among the extinct synapses undergo in-situ regeneration.
[0022] The initial activation characteristics and lifespan of all regenerated synapses are initialized to 0, and the dynamic balance between synapse death and regeneration is maintained continuously.
[0023] Secondly, the present invention provides a new energy output prediction system, comprising:
[0024] The data acquisition module is responsible for acquiring historical new energy output data with timestamps;
[0025] The new energy output prediction module is responsible for inputting historical new energy output data with timestamps into the trained, validated, and tested dynamic equilibrium time series network model to obtain new energy output prediction data.
[0026] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the method.
[0027] Fourthly, the present invention provides a machine-readable storage medium storing machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the method.
[0028] The present invention has the following advantages due to the adoption of the above technical solutions:
[0029] This invention achieves efficient pruning and regeneration of time-series networks by simulating the structural plasticity mechanism of biological nervous systems, forming a dynamically balanced network structure while maintaining the basic performance of the model, and even slightly improving it in some cases. It effectively adapts to the dynamic and random nature of new energy power output data, and improves the robustness and prediction accuracy of the model in actual power prediction tasks.
[0030] This invention comprehensively considers synaptic activation characteristics, survival time information, and random factors in pruning decisions, integrating short-term contribution, long-term stability, and exploratory nature, thus avoiding the limitations of single-indicator decisions. Activation characteristics reflect the degree of contribution of synapses to the current task, survival time information protects stable connections that survive long-term, and random factors increase the diversity of network structure. The combination of these three factors achieves a more scientific and rational adjustment of network structure.
[0031] The dynamic structure optimization method for predicting new energy output based on structural plasticity designed in this invention has good versatility and can be applied to various network architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and CNN-LSTM. Attached Figure Description
[0032] Figure 1 This is a flowchart of a new energy output prediction method based on structural plasticity provided in an embodiment of the present invention.
[0033] Figure 2 The results are the RMSE metrics for different models, where Sparse Models represent sparse CNN-LSTM models (i.e., the dynamic equilibrium temporal network model of this invention), and Original Model represents the original CNN-LSTM model (i.e., the model without sparsity).
[0034] Figure 3 These are the MAE index results applied using different models.
[0035] Figure 4 It is R that applies different models 2 Indicator results. Detailed Implementation
[0036] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.
[0037] A new energy output prediction method based on structural plasticity, see appendix. Figure 1 The method includes:
[0038] Obtain historical, timestamped renewable energy output data and manually annotate it;
[0039] The power data is preprocessed and then divided into training, testing, and validation sets according to a certain ratio.
[0040] A dynamic equilibrium time series network model is constructed and trained, validated, and tested using training, testing, and validation sets. The input of the dynamic equilibrium time series network model is historical renewable energy output data with timestamps, and the output is the predicted renewable energy output data for the future.
[0041] A well-trained, validated, and tested dynamic equilibrium time series network model is used for predicting the output of new energy sources.
[0042] The dynamically balanced temporal network model refers to the method by which neurons in existing temporal network models maintain network dynamic stability through the dynamic death and growth characteristics of synapses; specifically:
[0043] S1. Calculate the activation characteristic value based on the synaptic gradient, reflecting the degree of influence of the synapse on the model performance;
[0044] S2. Calculate the survival time of each synapse and construct a survival time control mechanism based on the survival time of the synapses;
[0045] S3. Calculate the probability of synapse extinction based on activation characteristic values and survival time control mechanisms;
[0046] S4. The extinction and regeneration of synapses are dynamically adjusted according to the extinction probability and activation characteristic value, and the dynamic stability of the network is maintained through the dynamic extinction and growth characteristics of synapses.
[0047] For example, the new energy power data includes at least one of the following: wind speed, load, oil temperature, and electric field output of the power equipment.
[0048] For example, step S1 simulates the time-cumulative process of synaptic changes. The calculated activation characteristics record the activity history of the synapse, and their calculation uses the following formula:
[0049] Equation (1)
[0050] in, and These are the first and second versions of the previous version. The activation characteristic value of the first synapse and the updated value of the second synapse. The activation characteristic value of a synapse. N represents the number of synapses. It is the EMA (Exponential Moving Average) smoothing coefficient. It is the absolute value of the current synaptic gradient.
[0051] For example, the survival time and survival time control mechanism are calculated using the following formula:
[0052] Equation (2)
[0053] Equation (3)
[0054] in, and They represent the first The survival time of the first synapse before its update, the first Lifetime of an updated synapse; It is a mask used to control whether the synapse lifespan needs to be updated; It is the first The survival time control value of a synapse; the longer the survival time, the smaller the value. It is the survival time constant, which controls the degree of influence of survival time on the survival time control value.
[0055] For example, the probability of extinction is calculated using the following formula:
[0056] Equation (4)
[0057] Equation (5)
[0058] in, It is the first after standardization The activation characteristic value of a synapse. It is the first The activation attribute value after the synapse is updated. This is the minimum value for activating the feature after the update. This is the maximum value activated after the update. It is the first The probability of synapse demise. It is the extinction coefficient, which scales the extinction probability. It is an exponential parameter. It is the first The survival time control value of each synapse.
[0059] For example, the process of dynamically adjusting synaptic extinction and regeneration based on extinction probability and activation characteristic value includes:
[0060] (1) Each synapse dies with a probability of demise, and the probability of demise is positively correlated with the synapse demise.
[0061] (2) After the synapse dies, the required number of regenerated synapses is obtained based on the target sparsity. The K synapses with the highest activation characteristic values among the extinct synapses are regenerated in situ, where 0 < K < 0. , ( The required regenerated synapses are regenerated using a random growth strategy to maintain network dynamic stability. This indicates the preset random growth ratio.
[0062] K= Equation (6)
[0063] (3) The initial activation characteristics and lifespan of each regenerated synapse are initialized to 0.
[0064] (4) Synaptic extinction and regeneration continue throughout the entire network training process, dynamically maintaining the sparse balance of the network.
[0065] For example, the existing temporal network model employs one of LSTM, GRU, or CNN-LSTM.
[0066] The following implementation case uses a CNN-LSTM model to demonstrate the specific implementation of a dynamic optimization method for predicting new energy output based on structural plasticity.
[0067] This implementation case uses a photovoltaic power output dataset from a certain region to conduct experiments, predicting power output data for the next hour. The results of the baseline model and the experimental results incorporating the method of this invention are as follows: Figures 2 to 4 This includes RMSE, MAE, and R, which contain sparsity values of 90% to 10%. 2 Indicators such as...
[0068] from Figures 2 to 4 As can be seen, the structural optimization method of the present invention outperforms the original model in multiple sparsity levels, indicating that the present invention can achieve structural optimization of the new energy output prediction network model, further improve prediction accuracy, and reduce network size.
Claims
1. A method for predicting the output of new energy sources based on structural plasticity, the method comprising: Obtain historical, timestamped renewable energy output data and manually annotate it; The power data is preprocessed and then divided into training set, validation set and test set according to the proportions. A dynamic equilibrium time series network model is constructed, and it is trained, validated, and tested using training, validation, and test sets. A well-trained, validated, and tested dynamic equilibrium time series network model is used for predicting the output of new energy sources. Its characteristic is that the dynamically balanced temporal network model maintains the dynamic stability of the network by utilizing the dynamic death and growth characteristics of synapses in the neurons of existing temporal network models; specifically: Activation characteristic values are calculated based on synaptic gradients; Calculate the survival time of each synapse, and construct a survival time control mechanism based on the synapse survival time; The probability of synapse extinction is calculated based on activation characteristic values and survival time control mechanisms; Synaptic extinction and regeneration are dynamically adjusted based on extinction probability and activation characteristic values.
2. The method according to claim 1, characterized in that, The new energy power data includes at least one of the following: wind speed, load, oil temperature, and electric field output of the power equipment.
3. The method according to claim 1, characterized in that, The activation characteristic value is calculated using the following formula: Equation (1) in, and These are the first and second versions of the previous version. The activation characteristic value of the first synapse and the updated value of the second synapse. The activation characteristic value of a synapse. N represents the number of synapses. It is the EMA smoothing coefficient. It is the absolute value of the current synaptic gradient.
4. The method according to claim 1, characterized in that, The survival time and survival time control mechanism are calculated using the following formula: Equation (2) Equation (3) in, and They represent the first The survival time of the first synapse before its update, the first Lifetime of an updated synapse; It's a mask; It is the first Survival time control values for each synapse; It is the survival time constant.
5. The method according to claim 1, characterized in that, The probability of extinction is calculated using the following formula: Equation (4) in, It is the first after standardization The activation characteristic value of a synapse. It is the first The activation attribute value after the synapse is updated. This is the minimum value for activating the feature after the update. This is the maximum value activated after the update. It is the first The probability of synapse demise. It is the extinction coefficient. It is an exponential parameter. It is the first The survival time control value of each synapse.
6. The method according to claim 1, characterized in that, The process of dynamically adjusting synaptic extinction and regeneration based on extinction probability and activation characteristic value includes: Each synapse dies with a probability of demise, and the probability of demise is positively correlated with the synapse demise. After the synapse dies, the required number of regenerated synapses is obtained based on the target sparsity. ; ( The required regenerated synapses regenerate new synapses through a random growth strategy. The growth ratio is random; the remaining number of synapses with the highest activation value among the extinct synapses undergo in-situ regeneration. The initial activation characteristics and lifespan of all regenerated synapses are initialized to 0, and the dynamic balance between synapse death and regeneration is maintained continuously.
7. The method according to claim 1, characterized in that, The existing temporal network model uses one of LSTM, GRU, or CNN-LSTM.
8. A new energy output prediction system implementing the method of any one of claims 1-7, characterized in that, include: The data acquisition module is responsible for acquiring historical new energy output data with timestamps; The new energy output prediction module is responsible for inputting historical new energy output data with timestamps into the trained, validated, and tested dynamic equilibrium time series network model to obtain new energy output prediction data.
9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the method of any one of claims 1-7.
10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the method of any one of claims 1-7.