A smart park load prediction method fusing seasonal solstice features

Abstract

This invention discloses a smart park load forecasting method integrating seasonal and solar term features, comprising: generating a feature input vector and a meteorological-solar term deviation vector based on current input data; a trained weight calculation module outputting a weight vector and a bidirectional dynamic hidden state vector based on these two vectors; a trained gating network outputting a fuzzy probability distribution based on the weight vector and the bidirectional dynamic hidden state vector; and a trained fuzzy boundary hybrid expert network outputting park load forecast data for a preset future time period based on the weight vector, the bidirectional dynamic hidden state vector, and the fuzzy probability distribution. The expert network includes multiple trained seasonal and solar term sub-networks, each of which uses the knowledge learned during training to predict park load data. This invention can deeply capture the nonlinear spatiotemporal relationships of multi-source data, significantly improving the prediction accuracy, convergence speed, and robustness of the model under complex and variable climate conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent forecasting technology, specifically relating to a smart park load forecasting method that integrates seasonal solar term characteristics. Background Technology

[0002] Among existing energy load forecasting technologies, deep learning-based load forecasting methods can improve forecast accuracy to some extent, but significant bottlenecks remain in terms of adaptability to atypical climates, dynamic feature coupling capabilities, and model switching smoothness. Current technologies often mechanically rely on calendar times (such as solar terms and months) as input, ignoring the "spatiotemporal misalignment" between actual meteorological seasons and calendar solar terms (such as late spring cold snaps and warm winters), leading to inaccurate predictions during climate anomalies. Furthermore, traditional feature engineering often uses static weight allocation, failing to effectively distinguish the dynamic drift of feature importance between stable and abrupt periods; while conventional hybrid expert networks (MoE) typically perform "hard partitioning" during seasonal transitions, easily creating prediction gaps and failing to meet the stringent requirements of smart parks for high timeliness and stability in energy management.

[0003] In other words, existing energy load forecasting technologies suffer from problems such as insufficient feature weight allocation, inadequate capture of nonlinear relationships, poor model seasonality generalization, and poor robustness, making it difficult to meet the requirements for high timeliness and high reliability in smart park energy management. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a smart park load forecasting method that integrates seasonal solar term characteristics. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a smart park load forecasting method that integrates seasonal and solar term characteristics, comprising: A feature input vector is generated based on the current input data, and a meteorological-solar term deviation vector is generated based on the current meteorological data and current time data with the current solar term label contained in the current input data, as well as the pre-built meteorological-solar term benchmark library. The feature input vector and the meteorological-solar term deviation vector are input into the trained weight calculation module, and the trained weight calculation module outputs the weight vector and the bidirectional dynamic hidden state vector. The weight vector and the bidirectional dynamic hidden state vector are input into the trained gating network, and the trained gating network outputs a fuzzy probability distribution. The weight vector, the bidirectional dynamic hidden state vector, and the fuzzy probability distribution are input into a trained fuzzy boundary hybrid expert network. The trained fuzzy boundary hybrid expert network outputs an expert network weight and output, which are used as the park load prediction data for a future preset time period. The trained fuzzy boundary hybrid expert network includes multiple trained seasonal solar term sub-networks. Each trained seasonal solar term sub-network uses the park load data for the corresponding solar term within the corresponding season learned during the training process, as well as the weight vector and the bidirectional dynamic hidden state vector, to predict the park load data. The fuzzy probability distribution contains the weights of the prediction results output by each seasonal solar term sub-network.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Overcoming the limitations of "calendar solar terms," ​​this invention significantly improves the robustness of predictions under atypical climate conditions: Compared to existing technologies that only use solar terms as time labels, this invention introduces the "meteorological-solar term deviation vector" feature. By calculating the deviation vector between real-time weather and the solar term benchmark, the model can keenly perceive abnormal climate events such as "late spring cold snaps" and "warm winters." This allows the prediction system to no longer be limited to a fixed calendar cycle, but to adaptively adjust according to the actual "meteorological season," effectively solving the problem of prediction inaccuracies caused by climate anomalies.

[0006] 2) This invention achieves "active denoising" and "dynamic focusing" of feature weights, solving the problem of poor adaptability of static feature engineering: Through a threshold truncation mechanism, this invention actively removes redundant features with low mutual information, preventing noisy data from interfering with model training. Furthermore, by coupling temperature scaling with the dynamic state of BiLSTM, this invention enables the model to focus on global patterns during periods of stable load, while rapidly "focusing" on key meteorological factors (such as temperature characteristics during sudden temperature drops) during periods of abrupt changes. This dynamic-static combined mechanism significantly improves the model's generalization ability under complex and variable operating conditions.

[0007] 3) Eliminating "prediction gaps" during seasonal transitions and ensuring smooth and stable prediction curves throughout the entire timeframe: Unlike the "rigid division of labor" in traditional MoE architectures, this invention endows the seasonal expert network with "fuzzy inference capabilities during transition periods." During seasonal transitions, the fuzzy probability distribution output by the gating network can control multiple seasonal expert subnetworks (i.e., multiple seasonal solar term subnetworks) to work in parallel at a specific ratio, achieving soft switching of the model's internal state. This not only avoids the risk of a sudden performance drop in a single model during seasonal changes but also significantly improves prediction stability and accuracy over long periods of operation.

[0008] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the adaptive energy load forecasting method that integrates meteorological deviation and dual-flow gating mechanism provided in this embodiment of the invention. Figure 2 This is a schematic diagram illustrating an exemplary reasoning and training process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of a smart park load forecasting system that integrates seasonal solar term characteristics, provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0011] This invention designs an adaptive energy load forecasting method that integrates meteorological bias and a dual-stream gating mechanism. This method breaks through the limitation of a single time label by constructing a "meteorological-seasonal bias vector" to transform discrete seasonal information into continuous real-time climate features. It abandons traditional static weighting and designs a dual-stream spatiotemporal coupling gating mechanism combining mutual information (MI) and BiLSTM states, utilizing temperature scaling technology to achieve dynamic focusing and denoising of multi-source features. Furthermore, a fuzzy inference mechanism is introduced within a hybrid expert network (MoE) framework to control the soft switching and parallel collaboration of the seasonal expert subnet. During training, an RMSprop optimization strategy is employed to improve convergence performance. Through these designs, this invention can deeply capture the nonlinear spatiotemporal relationships of multi-source data, significantly improving the prediction accuracy, convergence speed, and robustness of the model under complex and variable climate conditions.

[0012] Figure 1 This is a flowchart illustrating a smart park load forecasting method that integrates seasonal solar term characteristics, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Generate a feature input vector based on the current input data, and generate a meteorological-solar term deviation vector based on the current meteorological data and current time data with the current solar term label contained in the current input data, as well as the pre-built meteorological-solar term benchmark library.

[0013] Here, the current input data includes: current energy unit price (e.g., electricity price, gas price, coal price), current meteorological data, and current time data with the current solar term tag. It should be noted that "current" in this invention refers to a current period of time, such as the current 15 minutes, the current day (i.e., today), etc. Meteorological data includes temperature, humidity, wind speed, irradiance, etc. Current time data specifically refers to time data such as year, month, and day. The solar term tag accompanying the current time data indicates the solar term in which the current time data is located. The meteorological-solar term benchmark database is constructed based on collected historical meteorological data and contains standard meteorological data for each solar term in each season of each year.

[0014] S102. Input the feature input vector and the meteorological-solar term deviation vector into the trained weight calculation module, and the trained weight calculation module outputs the weight vector and the bidirectional dynamic hidden state vector.

[0015] Here, the weight calculation module includes a static flow weight calculation layer and a dynamic flow weight calculation layer, which operate in parallel. This invention, by designing a weight calculation module that includes both static and dynamic flow weight calculation layers, abandons the single feature input method, constructs parallel processing of static and dynamic flows, and achieves feature enhancement through an improved weighting mechanism.

[0016] S103. Input the weight vector and the bidirectional dynamic hidden state vector into the trained gating network, and the trained gating network outputs a fuzzy probability distribution. It should be noted that the fuzzy probability distribution is the weight corresponding to the prediction result output by the solar term sub-network for each season.

[0017] For example, a gating network is a nonlinear mapping layer. The trained gating network simultaneously receives weight vectors and bidirectional dynamic hidden state vectors. Through the nonlinear mapping layer, static and dynamic features are mapped to the same high-dimensional space for interaction, and a fuzzy probability distribution that reflects the "seasonal membership degree" of the current time data is calculated.

[0018] S104. Input the weight vector, bidirectional dynamic hidden state vector, and fuzzy probability distribution into the trained fuzzy boundary hybrid expert network. The trained fuzzy boundary hybrid expert network outputs the expert network weights and output, which are used as the park load prediction data for a future preset time period. The trained fuzzy boundary hybrid expert network includes: multiple trained seasonal solar term sub-networks. Each trained seasonal solar term sub-network is used to predict the park load data using the park load data for the corresponding solar term in the corresponding season learned during the training process, as well as the weight vector and bidirectional dynamic hidden state vector.

[0019] For example, the trained fuzzy boundary fusion expert network includes: a trained spring solar term sub-network, a trained summer solar term sub-network, a trained autumn solar term sub-network, and a trained winter solar term sub-network. The trained summer, autumn, and winter solar term sub-networks constitute four expert sub-networks or four seasonal expert sub-networks. Each seasonal solar term sub-network can be a BiLSTM network. The trained spring, summer, autumn, and winter solar term sub-networks operate in parallel.

[0020] It should be noted that the aforementioned "preset time period in the future" can be the next 15 minutes or the next day, etc., and is specifically determined by the relevant parameters during the training process. This invention does not limit this.

[0021] It should be noted that the park load data refers to the park's energy load data. For example, when the energy source is electricity, the park load data refers to the park's electricity load data (i.e., electricity consumption data). Similarly, when the energy source is natural gas, the park load data refers to the park's natural gas load data (i.e., natural gas consumption data).

[0022] In some embodiments, S101 is implemented through steps S1011 to S1013: S1011. Construct a vector based on the current energy price, current meteorological data, and current time data with the current solar term label to obtain the feature input vector.

[0023] S1012. Based on the current solar term label, find the standard meteorological data corresponding to the current solar term label from the meteorological-solar term benchmark database. The meteorological-solar term benchmark database contains standard meteorological data for each solar term in each season.

[0024] S1013. Construct a difference vector based on the difference between the current meteorological data and the standard meteorological data corresponding to the current solar term label to obtain the meteorological-solar term deviation vector.

[0025] For example, when meteorological data includes temperature, humidity, wind speed, and irradiance, the meteorological-solar term deviation vector is a vector composed of the temperature difference, humidity difference, wind speed difference, and irradiance difference between the current temperature, humidity, wind speed, and irradiance and the standard temperature, humidity, wind speed, and irradiance found from the meteorological-solar term reference library. In this way, the present invention can transform traditional discrete-time features into a continuous "meteorological-solar term deviation vector." This step solves the deficiency that using only calendar solar terms cannot reflect the impact of abnormal climate (such as "late spring cold snap" or "warm winter") on the load.

[0026] In some embodiments, S102 is implemented through steps S1021 to S1022: S1021. Input the feature input vector and the meteorological-solar term deviation vector into the trained static flow weight calculation layer. The trained static flow weight calculation layer generates a weight vector and outputs it based on mutual information analysis and truncated temperature scaling Softmax mechanism.

[0027] Specifically, S1021 is achieved through steps S10 to S14: S10. Input the feature input vector and the meteorological-seasonal deviation vector into the trained static flow weight calculation layer. The trained static flow weight calculation layer uses the mutual information analysis method to calculate the correlation between each element in the feature input vector and the meteorological-seasonal deviation vector and the park load data, and obtains the mutual information score of each element in the feature input vector and the meteorological-seasonal deviation vector. .

[0028] It should be noted that the use of mutual information analysis to calculate the correlation between each element in the feature input vector and the meteorological-solar term deviation vector and the park load data is an existing technology, and this invention does not limit it.

[0029] S20. The trained static flow weight calculation layer utilizes a preset threshold δ and mutual information score. The feature input vector is denoised to obtain the denoised feature input vector.

[0030] Here, the preset threshold δ can be pre-set according to actual needs, and this invention does not specifically limit its value. Specifically, the mutual information score in the feature input vector... The values ​​of elements smaller than the preset threshold δ are all set to 0. In this way, the model can be prevented from being disturbed by noisy data while keeping the data dimension unchanged.

[0031] S30. The trained static flow weight calculation layer normalizes the temperature contained in the feature input vector and the meteorological-solar term deviation vector to obtain the temperature coefficient. Using temperature coefficient The mutual information score of each element in the denoised feature input vector and the meteorological-solar term deviation vector. Adjustments are made to obtain the adjusted mutual information score. .

[0032] Specifically, the formula for calculating the adjusted mutual information score of any element in the denoised feature input vector and the meteorological-solar term deviation vector is as follows: ; in, It is the adjusted mutual information score of any one of the elements. It is the mutual information score of any element.

[0033] Here, the temperature data is normalized to obtain the temperature coefficient. Distributed between 0 and 2, utilizing temperature coefficient The entropy of the weight distribution can be controlled. Specifically, when When <1, the distribution can be sharpened, forcing the model to focus only on the most correlated "head features" (suitable for extreme weather); when When the value is greater than 1, the distribution is smooth and more "long-tail feature" information is retained.

[0034] S40. The mutual information score of the trained static flow weight calculation layer after adjusting each element in the denoised feature input vector and the meteorological-solar term deviation vector. Softmax processing is performed to obtain a static weight vector that is adaptable to operating conditions and reflects long-term patterns. The weight vector contains the weights of each element in the denoised feature input vector and the meteorological-solar term deviation vector.

[0035] S1022. Input the feature input vector and the meteorological-solar term deviation vector into the trained dynamic flow weight calculation layer. The trained dynamic flow weight calculation layer generates a bidirectional dynamic hidden state vector and outputs it.

[0036] For example, the dynamic flow weight calculation layer is a BiLSTM network. This BiLSTM network can be used to process the original multi-source sequences to capture the historical dependencies and future trend constraints of the time series bidirectionally, and finally output a dynamic hidden state containing deep temporal patterns (i.e., a bidirectional dynamic hidden state vector). It should be noted that this bidirectional dynamic hidden state vector is a dynamic weight vector.

[0037] In some embodiments, the above-mentioned S104 is implemented through steps S1041 to S1042: S1041. Input the weight vector, bidirectional dynamic hidden state vector, and fuzzy probability distribution into the trained fuzzy boundary hybrid expert network. The trained seasonal solar term subnetworks in the trained fuzzy boundary hybrid expert network predict the park load data based on the weight vector and bidirectional dynamic hidden state vector, and output the prediction results.

[0038] S1042. The fuzzy boundary hybrid expert network, based on the fuzzy probability distribution, performs a weighted summation of the prediction results of multiple trained seasonal solar term sub-networks to obtain the expert network weighted output, which is then used as the park load prediction data for a future preset time period.

[0039] Specifically, the trained spring, summer, autumn, and winter sub-networks predict park load data based on weight vectors and bidirectional dynamic hidden state vectors, and output corresponding prediction results. Then, using the weights of these sub-networks contained in the fuzzy probability distribution, the prediction results are weighted and summed to obtain the expert network weighted output. This expert network weighted output is the park load prediction data for a predetermined time period. Clearly, in this invention, while each expert sub-network focuses on fitting the nonlinear load pattern under a specific seasonal model, a fuzzy gating mechanism allows each expert sub-network to be activated across seasons during atypical seasons (such as "late spring cold snaps"), achieving soft switching processing for complex operating conditions and significantly improving the model's generalization ability under boundary conditions. Furthermore, this invention uses a unified hybrid expert network, combined with a gating mechanism, to dynamically select and fuse experts from the four seasons within the same framework. This not only reduces training and maintenance costs but also better captures seasonal differences, thereby improving overall prediction accuracy and robustness.

[0040] In this invention, a trained weight calculation module, a trained gating network, and a trained fuzzy boundary hybrid expert network constitute a trained load forecasting model. The training method for this trained load forecasting model includes: S20. Obtain historical data for a preset historical time period, wherein the preset historical time period includes each solar term of each season, and the historical data for the preset historical time period includes: park load data for the preset historical time period, electricity price for the preset historical time period, time data with solar term labels for the preset historical time period, and meteorological data for the preset historical time period.

[0041] It should be noted that the preset historical time period can be set according to actual needs, such as one year or two years, but it must cover every solar term in every season of the year. For example, when the preset historical time period is one year, the historical data for that year includes: the park's load data for that year, the energy price for that year, the date and corresponding solar term for that year, and the meteorological data for that year.

[0042] S21. Based on historical data for a preset historical time period, construct a dataset containing multiple training samples and construct an initial load prediction model. Each training sample contains a sample feature input vector and a meteorological-solar term deviation vector. Each sample feature input vector is a vector constructed based on the park load data, energy prices, time data with solar term labels, and meteorological data for a preset historical time period. The label of each training sample is the actual park load data for the preset time period.

[0043] As mentioned above, the preset time period can be 15 minutes or the next day, etc. Specifically, for example, when the preset historical time period is one year and the preset time period is one day, after obtaining one year of historical data, the historical data of this year is divided into 365 days of historical data, and the historical data of each day is used as a training sample, and the actual load data of the park each day is used as the label of the training sample. In this way, each training sample includes: the actual park load data of this day, the energy price of this day, the weather data of this day, the year, month and day data of this day and the corresponding solar term.

[0044] S22. Each time the load prediction model is trained, the training samples for this training session are obtained from the dataset and input into the load prediction model to be trained. The expert network weights and outputs for this training session are output. The loss value for this training session is calculated based on the labels of the training samples and the expert network weights and outputs for this training session. The loss value for this training session is used for backpropagation to adjust the network parameters of the weight calculation module, the gating network, and the fuzzy boundary hybrid expert network to obtain the load prediction model trained for this training session. This process is iterated until the number of training sessions meets the threshold or the loss value reaches the preset condition, and the trained load prediction model is obtained.

[0045] For example, the loss function during training can be the MSE / RMSE function. Furthermore, during training, the RMSprop optimizer can be used for parameter updates to accommodate the non-stationary gradient characteristics.

[0046] It should be noted that the preset conditions can be set according to actual needs. For example, the obtained loss value may be less than a preset loss threshold, or the difference between two consecutive loss values ​​may be less than a preset difference. This invention does not limit these conditions.

[0047] For example, Figure 2 This is a schematic diagram of a reasoning and training process provided by the present invention. For example... Figure 2As shown, the black arrows represent the flowcharts in the inference process (i.e., the actual usage process), and the red arrows represent the flowcharts in the training process. Here, (MI+Softmax)* represents the static flow weight calculation layer, and BiLSTM Layer represents the dynamic flow weight calculation layer.

[0048] This invention also provides a smart park load forecasting system that integrates seasonal and solar term characteristics, such as... Figure 3 As shown, the system includes: a data acquisition terminal, a model training and prediction platform, and embedded devices. The data acquisition terminal collects raw data required for the training or inference process, such as energy prices, meteorological data, time data, and corresponding solar term labels. The model training and prediction platform is used to implement the training method for the load prediction model, thereby completing the training of the load prediction model and obtaining a trained load prediction model. The embedded devices are used to carry the trained load prediction model and receive data collected by the data acquisition terminal in real time to predict the park's load data. The obtained prediction results can further support the park's energy dispatching.

[0049] In summary, compared with the prior art, the present invention includes the following core features: (1) A dynamic feature enhancement method based on the “meteorological-solar term deviation vector” is proposed. Specifically, unlike the traditional approach of only using solar terms as discrete time labels, this invention constructs a meteorological-solar term benchmark library. By calculating the deviation vector between real-time meteorological data (temperature, humidity, irradiance, etc.) and the historical average of the current solar term, this deviation vector is used as a correction factor to solve the problem of model inaccuracy caused by the asynchronous relationship between “calendar solar terms” and “actual meteorological season” (such as “late spring cold” or “warm winter”) in the traditional method, and significantly improves the robustness of load prediction under atypical climatic conditions.

[0050] (2) This invention designs a "dual-stream attention collaborative gating mechanism" integrating "threshold truncation and temperature scaling". Specifically, addressing the problem of weak generalization ability and susceptibility to noise interference in traditional Softmax gating, this invention proposes an improved dual-stream architecture: the static stream introduces a mutual information threshold truncation mechanism to actively eliminate interference from weakly correlated features, and uses temperature scaling technology to dynamically adjust the distribution sharpness of feature weights, enabling the model to "force focus" on core influencing factors under extreme weather conditions; the dynamic stream uses BiLSTM to extract bidirectional dynamic hidden state vectors to capture the temporal features of load mutations. The two are fused nonlinearly to generate the final activation coefficients. This mechanism not only achieves active noise reduction at the feature level, but also takes into account the dual constraints of statistical regularity and real-time fluctuations, significantly improving the feature capture sensitivity of the model under complex working conditions.

[0051] (3) A seasonal hybrid expert network with "transitional period fuzzy inference capability" was constructed: Based on the traditional MoE architecture, this invention introduces a "transitional period fuzzy response mechanism" to address the meteorological complexity during seasonal transitions. The gated network can identify seasonal transition characteristics (such as the transition between spring and summer) and control the corresponding two seasonal expert subnetworks (such as the spring subnetwork and the summer subnetwork) to work in parallel at a specific ratio, rather than a simple binary switching. This soft switching mechanism effectively eliminates the prediction gaps caused by rigid seasonal divisions and improves the smoothness and accuracy of the model during seasonal transitions.

[0052] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0053] In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. While different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0054] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A smart park load forecasting method integrating seasonal solar term characteristics, characterized in that, include: A feature input vector is generated based on the current input data, and a meteorological-solar term deviation vector is generated based on the current meteorological data and current time data with the current solar term label contained in the current input data, as well as the pre-built meteorological-solar term benchmark library. The feature input vector and the meteorological-solar term deviation vector are input into the trained weight calculation module, and the trained weight calculation module outputs the weight vector and the bidirectional dynamic hidden state vector. The weight vector and the bidirectional dynamic hidden state vector are input into the trained gating network, and the trained gating network outputs a fuzzy probability distribution. The weight vector, the bidirectional dynamic hidden state vector, and the fuzzy probability distribution are input into a trained fuzzy boundary hybrid expert network. The trained fuzzy boundary hybrid expert network outputs an expert network weight and output, which are used as the park load prediction data for a future preset time period. The trained fuzzy boundary hybrid expert network includes multiple trained seasonal solar term sub-networks. Each trained seasonal solar term sub-network uses the park load data for the corresponding solar term within the corresponding season learned during the training process, as well as the weight vector and the bidirectional dynamic hidden state vector, to predict the park load data. The fuzzy probability distribution contains the weights of the prediction results output by each seasonal solar term sub-network.

2. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, The step of generating a feature input vector based on the current input data, and generating a meteorological-solar term deviation vector based on the current meteorological data and current time data with the current solar term label contained in the current input data, as well as a pre-built meteorological-solar term benchmark library, includes: A vector is constructed based on current energy prices, current meteorological data, and current time data with current solar term labels to obtain the feature input vector; Based on the current solar term label, the standard meteorological data corresponding to the current solar term label is found from the meteorological-solar term benchmark library, wherein the meteorological-solar term benchmark library contains standard meteorological data for each solar term in each season; A difference vector is constructed based on the difference between the current meteorological data and the standard meteorological data corresponding to the current solar term label, thus obtaining the meteorological-solar term deviation vector.

3. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, The weight calculation module includes a static flow weight calculation layer and a dynamic flow weight calculation layer; the step of inputting the feature input vector and the meteorological-solar term deviation vector into the trained weight calculation module, and having the trained weight calculation module output a weight vector and a bidirectional dynamic hidden state vector, includes: The feature input vector and the meteorological-solar term deviation vector are input into the trained static flow weight calculation layer. The trained static flow weight calculation layer generates a weight vector and outputs it based on mutual information analysis and a truncated temperature scaling Softmax mechanism. The feature input vector and the meteorological-solar term deviation vector are input into the trained dynamic flow weight calculation layer, which generates and outputs a bidirectional dynamic hidden state vector.

4. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 3, characterized in that, Both the feature input vector and the meteorological-solar term deviation vector contain current temperature data; the feature input vector and the meteorological-solar term deviation vector are input into a trained static flow weight calculation layer, which generates and outputs a weight vector based on mutual information analysis and a truncated temperature scaling Softmax mechanism, including: The feature input vector and the meteorological-seasonal deviation vector are input into the trained static flow weight calculation layer. The trained static flow weight calculation layer uses mutual information analysis to calculate the correlation between each element in the feature input vector and the meteorological-seasonal deviation vector and the park load data, respectively, to obtain the mutual information score of each element in the feature input vector and the meteorological-seasonal deviation vector. ; The trained static flow weight calculation layer utilizes a preset threshold δ and the mutual information score. The feature input vector is denoised to obtain a denoised feature input vector. The trained static flow weight calculation layer normalizes the current temperature data to obtain the temperature coefficient. Using the temperature coefficient The mutual information score of each element in the denoised feature input vector and the meteorological-solar term deviation vector. Adjustments are made to obtain the adjusted mutual information score. ; The trained static flow weight calculation layer adjusts the mutual information score of each element in the denoised feature input vector and the meteorological-solar term deviation vector. Softmax processing is performed to obtain a weight vector, wherein the weight vector contains the weights of each element in the denoised feature input vector and the meteorological-solar term deviation vector.

5. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 4, characterized in that, The formula for calculating the adjusted mutual information score of any element in the denoised feature input vector and the meteorological-solar term deviation vector is as follows: ； in, It is the adjusted mutual information score of any one of the elements. It is the mutual information score of any one of the elements.

6. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 3, characterized in that, The dynamic flow weight calculation layer is a BiLSTM network.

7. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, The step of inputting the weight vector, the bidirectional dynamic hidden state vector, and the fuzzy probability distribution into a trained fuzzy boundary hybrid expert network, and outputting the expert network weights and output by the trained fuzzy boundary hybrid expert network, includes: The weight vector, the bidirectional dynamic hidden state vector, and the fuzzy probability distribution are input into the trained fuzzy boundary hybrid expert network. Each seasonal solar term subnetwork in the trained fuzzy boundary hybrid expert network predicts the park load data based on the weight vector and the bidirectional dynamic hidden state vector, and outputs the prediction results. The fuzzy boundary hybrid expert network performs a weighted summation of the prediction results of the trained multiple seasonal solar term sub-networks based on the fuzzy probability distribution, and obtains the expert network weighted output, which is then used as the park load prediction data for a future preset time period.

8. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, Each seasonal solar term subnetwork is a BiLSTM network.

9. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, The trained fuzzy boundary hybrid expert network includes: a trained spring seasonal sub-network, a trained summer seasonal sub-network, a trained autumn seasonal sub-network, and a trained winter seasonal sub-network.

10. The smart park load forecasting method integrating seasonal solar term characteristics according to claim 1, characterized in that, The trained weight calculation module, the trained gating network, and the trained fuzzy boundary hybrid expert network constitute the trained load prediction model. The training method of the trained load prediction model includes: Acquire historical data for a preset historical time period, wherein the preset historical time period includes each solar term of each season, and the historical data for the preset historical time period includes: park load data for the preset historical time period, energy prices for the preset historical time period, time data with solar term tags for the preset historical time period, and meteorological data for the preset historical time period. Based on historical data from the preset historical time period, a dataset containing multiple training samples is constructed, and an initial load prediction model is built. Each training sample contains a sample feature input vector and a meteorological-solar term deviation vector. Each sample feature input vector is a vector constructed based on the park load data, energy prices, time data with solar term labels, and meteorological data of the preset historical time period. The label of each training sample is the actual park load data of the preset time period. Each time the load prediction model is trained, the training samples for this training session are obtained from the dataset and input into the load prediction model to be trained. The expert network weights and outputs for this training session are output. The loss value for this training session is calculated based on the labels of the training samples and the expert network weights and outputs for this training session. Backward gradient propagation is performed using the loss value to adjust the network parameters of the weight calculation module, the gating network, and the fuzzy boundary hybrid expert network to obtain the load prediction model trained for this training session. This process is iterated until the number of training sessions meets the threshold or the loss value reaches the preset condition, thus obtaining the trained load prediction model.

Images