A lightweight dynamic air conditioning load regulation capacity prediction method and system with physical information embedding

Abstract

This invention relates to a lightweight dynamic air conditioning load regulation capability prediction method and system based on physical information embedding. The method includes: collecting and preprocessing multi-source heterogeneous data required for prediction; using the preprocessed multi-source heterogeneous data, based on physical information constraints, and through a pre-constructed lightweight dynamic hybrid model, outputting the baseline value and dynamic change trend of air conditioning load for future periods; calculating the adjustable power range, regulation rate, and duration based on the baseline value and dynamic change trend of air conditioning load for future periods, providing quantitative basis for power grid dispatch; continuously monitoring the prediction error of the lightweight dynamic hybrid model, and triggering an online learning process to adapt to data distribution drift when the error exceeds a preset threshold or reaches a set update cycle. Compared with existing technologies, this invention achieves rapid and reliable assessment of regulation capability in new scenarios, improving the practicality and promotional value of prediction methods.

Description

Technical Field

[0001] This invention relates to the field of power system load regulation, and in particular to a lightweight dynamic air conditioning load regulation capability prediction method with embedded physical information. Background Technology

[0002] With the accelerated global energy transition and the deepening of the "dual carbon" goal, building a clean, low-carbon, safe, and efficient new power system has become an inevitable trend. In this process, the power grid faces the challenge of increasingly large peak-valley differences, especially during the high-temperature summer months when the concentrated start-up of air conditioning loads often leads to peak electricity consumption, placing enormous pressure on the safe and stable operation of the power grid. As a widely distributed, large-scale electrical device with thermal energy storage characteristics, air conditioning load has attracted widespread attention for its regulation potential. By scientifically predicting and effectively utilizing the regulation capabilities of air conditioning load, enabling it to proactively reduce or shift some electricity consumption during peak hours and moderately increase electricity consumption during off-peak hours, it is possible to effectively "shave peaks and fill valleys," smooth load fluctuations, and improve the economy and reliability of power grid operation.

[0003] Air conditioning loads possess a natural advantage as a high-quality, flexible regulation resource for the power grid. Their operation is closely related to temperature and exhibits a certain degree of thermal inertia; short-term start-ups, shutdowns, or power adjustments do not immediately and significantly impact indoor comfort, providing a window of opportunity for grid interaction. Accurately assessing the adjustability potential of air conditioning loads is a crucial prerequisite for transforming them from "passive electrical equipment" to "active grid resources." Accurately predicting future trends in air conditioning load and their adjustability boundaries helps power grid dispatching departments develop scientific load management strategies in advance, maximizing the regulatory value of demand-side resources while ensuring basic user comfort.

[0004] Integrating air conditioning load into the power grid regulation system not only helps alleviate peak power shortages and reduce reliance on new power generation capacity, but also promotes the consumption of renewable energy. When wind and solar power output is high, it can guide air conditioning load to increase electricity consumption, and vice versa, thereby enhancing the power system's ability to cope with uncertainties on both the power source and load sides. Therefore, in-depth research into the mechanism and potential assessment methods of air conditioning load in power grid peak-valley regulation is of significant practical importance for promoting the optimal allocation of energy and the sustainable development of the power system. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a lightweight dynamic air conditioning load regulation capability prediction method with physical information embedding. This solves the problem that traditional engineering methods are based on steady-state assumptions, which make it difficult to handle rapidly changing dynamic processes. The model training and prediction calculations are also time-consuming and cannot meet the requirements of real-time updates.

[0006] The objective of this invention can be achieved through the following technical solutions: A lightweight dynamic air conditioning load regulation capability prediction method embedding physical information, the method comprising: Collect and preprocess multi-source heterogeneous data required for prediction. The multi-source heterogeneous data includes outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting the thermal characteristics of buildings. Using the preprocessed multi-source heterogeneous data, and based on physical information constraints, a pre-constructed lightweight dynamic hybrid model is used to output the baseline value and dynamic trend of air conditioning load for future periods. Based on the baseline value and dynamic trend of air conditioning load in the future time period, the adjustable power range, adjustment rate and duration are calculated to provide a quantitative basis for power grid dispatch. The prediction error of the lightweight dynamic hybrid model is continuously monitored. When the error exceeds the preset threshold or reaches the set update cycle, the online learning process of the model is triggered to adapt to the data distribution drift.

[0007] Furthermore, the preprocessing procedure includes: The collected data is cleaned to remove noise and outliers; Missing values ​​are handled using interpolation methods; For data from different sources and at different frequencies, time alignment and feature extraction are performed to form a unified, high-quality feature vector.

[0008] Furthermore, the construction process of the lightweight dynamic hybrid model includes: A simplified first-order RC network model is established to describe the dynamic temperature changes in a single area of ​​a building. The first-order RC network model is solved to obtain the physical feature set. A physical neural network is constructed based on the physical feature set, and the physical feature set is used as part of the input features; A lightweight neural network model is established to characterize the dynamic response of air conditioning load. The output of the lightweight neural network model is the predicted indoor temperature. The RC thermal equilibrium differential equation is embedded as a physical constraint term into the loss function of the lightweight neural network model. The loss function is a weighted sum of data loss and physical constraint loss. With the goal of minimizing the loss function, the lightweight neural network model is trained using the loss function, and finally a pre-built lightweight dynamic hybrid model is obtained.

[0009] Furthermore, after training the lightweight neural network model, redundant components are removed and key structures are retained through weight / channel importance evaluation to construct a compact computational network; the high-precision floating-point parameters of the lightweight neural network model are converted into low-precision integers; and the types of network components are determined by the parameter reduction ratio.

[0010] Furthermore, the expression for the loss function is as follows: in, These are the physical constraint weighting coefficients. The number of training samples, For the first i The actual indoor temperature of each sample For the network prediction value, The equivalent heat capacity of the building's interior space. The equivalent thermal resistance of the building envelope. and Indoor and outdoor temperatures, respectively. The cooling or heating power provided to the air conditioning system.

[0011] Furthermore, the expression for the first-order RC network model is: in, The equivalent heat capacity of the building's interior space. The equivalent thermal resistance of the building envelope. and Indoor and outdoor temperatures, respectively. The cooling or heating power provided to the air conditioning system, Heat is gained internally.

[0012] Furthermore, the process of triggering the online learning process of the model to adapt to data distribution drift includes: Obtain the random variables of the load forecast output from the model, and represent the probability distribution of the model output using a Gaussian distribution; Based on the probability distribution output by the model, the confidence interval is calculated and obtained; Based on the confidence interval half-width and the reserved additional reserve capacity, obtain the actual executable effective reduction instructions; The sliding window of the dataset used to train the model is dynamically updated based on a preset sliding time window, and the lightweight dynamic hybrid model is retrained after the preset time period ends or the number of samples reaches a preset threshold; during full retraining, the model parameters are updated by minimizing the loss of the window dataset.

[0013] Furthermore, in the retraining process, to avoid the high computational cost of full retraining, incremental learning only uses the newly added sample set to update parameters, and a forgetting suppression regularization term is introduced.

[0014] A lightweight dynamic air conditioning load regulation capability prediction system with embedded physical information, the system comprising: a data input module, a preprocessing and fusion module, a hybrid model prediction module, an adaptive update module, and an output module; The data input module is responsible for collecting multi-source heterogeneous data required for prediction, mainly including outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting building thermal characteristics. The preprocessing and fusion module cleans the collected data to remove noise and outliers, and processes missing values ​​through interpolation. For data from different sources and at different frequencies, it performs time alignment and feature extraction to form a unified, high-quality feature vector. The hybrid model prediction module is an organically integrated lightweight dynamic hybrid architecture that outputs the basic value and dynamic trend of air conditioning load for future periods based on physical information constraints. The adaptive update module continuously monitors the model's prediction error. When the error exceeds a preset threshold or reaches the set update cycle, it triggers the model's online learning process to ensure that the model can adapt to data distribution drift caused by environmental changes, equipment aging, or changes in user behavior patterns, and maintain its long-term prediction accuracy. The output module calculates the adjustable power based on the predicted load dynamics.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the lightweight dynamic air conditioning load control capability prediction method with embedded physical information as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention include: This invention constructs a multi-source heterogeneous data fusion and enhancement framework, systematically integrating data from multiple dimensions such as outdoor environment, building structure, equipment operation, and user behavior. It establishes a lightweight dynamic hybrid model with embedded physical information, integrating known physical laws or prior knowledge into a data-driven machine learning framework in a constrained or guided manner, thereby improving model accuracy while enhancing its generalization ability and computational efficiency. An online adaptive learning mechanism for uncertainty is designed, aiming to enable the prediction model to continuously adapt to uncertainties brought about by environmental changes, equipment aging, and adjustments in user behavior patterns, ensuring its long-term robustness and reliability. An efficient "cold start" strategy is proposed, utilizing common knowledge across building and regional scenarios for transfer learning to achieve rapid and reliable evaluation of the ability to adapt to new scenarios, improving the practicality and promotional value of the prediction method. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram illustrating the functional modules of the system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] Example 1 This embodiment discloses a lightweight dynamic air conditioning load control capability prediction method embedding physical information, the method being as follows: Figure 1 As shown, it includes: Step S1: Collect and preprocess the multi-source heterogeneous data required for prediction. The multi-source heterogeneous data includes outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting the thermal characteristics of buildings.

[0020] The preprocessing process includes: The collected data is cleaned to remove noise and outliers; Missing values ​​are handled using interpolation methods; For data from different sources and at different frequencies, time alignment and feature extraction are performed to form a unified, high-quality feature vector.

[0021] Step S2: Using the preprocessed multi-source heterogeneous data and based on physical information constraints, output the basic value and dynamic trend of air conditioning load for future periods through a pre-built lightweight dynamic hybrid model.

[0022] The construction process of a lightweight dynamic hybrid model includes: A simplified first-order RC network model is established to describe the dynamic temperature changes in a single area of ​​a building. The first-order RC network model is solved to obtain the physical feature set. A physical neural network is constructed based on a physical feature set, which is used as part of the input features. To characterize the dynamic response of air conditioning load, a lightweight neural network model is established, and the output of the lightweight neural network model is the predicted indoor temperature. The RC thermal equilibrium differential equation is embedded as a physical constraint term into the loss function of the lightweight neural network model. The loss function is a weighted sum of data loss and physical constraint loss. With the goal of minimizing the loss function, a lightweight neural network model is trained using the loss function, ultimately obtaining a pre-built lightweight dynamic hybrid model.

[0023] After training the lightweight neural network model, redundant components are removed and key structures are retained through weight / channel importance evaluation to build a compact computational network; the high-precision floating-point parameters of the lightweight neural network model are converted into low-precision integers; and the types of network components are determined by the parameter reduction ratio.

[0024] Specifically, the first step is to establish a simplified first-order RC network model to describe the dynamic temperature changes in a single area of ​​the building. This model treats the interior and exterior of the building as two nodes, connected by an equivalent thermal resistance and an equivalent heat capacity, and represents the relationship between the rate of change of indoor temperature and the temperature difference between indoor and outdoor areas, as well as the cooling / heating power of the air conditioning system.

[0025] The expression for a first-order RC network model is: in, The equivalent heat capacity of the building's interior space. The equivalent thermal resistance of the building envelope. and Indoor and outdoor temperatures, respectively. The cooling or heating power provided to the air conditioning system, For internal heat gain (such as heat generated by people or equipment).

[0026] This formula indicates that the rate of change of indoor temperature is determined by the heat flow entering (or leaving) through the building envelope, the heat flow provided by air conditioning, and other internal heat sources. This formula abstracts the complex building heat transfer process into a calculable dynamic system, where heat capacity... The rate of temperature change (i.e., the magnitude of thermal inertia) and thermal resistance are determined by this. This affects the intensity of heat transfer.

[0027] Solving this first-order linear differential equation yields the time constant of the temperature change. τ (The core features of the RC model) are: τ = RC Therefore, the set of physical features extracted from the RC model can be represented as: The second step is to construct a physical neural network. Features derived from the RC model (such as the time constant of temperature change) can be used as part of the input features to guide the machine learning model to learn patterns that conform to physical laws. By combining a simplified RC network model with a data-driven model, a more reliable characterization of the dynamic response characteristics of air conditioning loads can be achieved while ensuring the model remains lightweight.

[0028] The physical prior features are fused with the features of the original monitoring data to form the total input vector of the prediction model: in, For time variables (used to calculate the derivative of temperature with respect to time, supporting the differential equations of physical constraints). (Please solve) For data-driven raw features, such as: in, For time characteristics (such as time of day, weekday / weekend).

[0029] The third step is to build a fusion model training network.

[0030] A lightweight neural network (MLP) model is established to characterize the dynamic response of air conditioning load. The output is the predicted indoor temperature. The network expression is: in, , , , , , , which are the weight matrices / vectors and biases for each layer; Lightweight activation functions (such as ReLU) σ ( z ) = max(0, z ), GELU σ ( z ) =z ⋅ Φ ( z (This reduces computational complexity).

[0031] The RC thermal equilibrium differential equation is used as a physical constraint term and embedded in the loss function to force the network output to satisfy physical laws. The total loss function is a weighted sum of data loss and physical constraint loss: in, Physical constraint weighting coefficients ( λ >0 can be adjusted through cross-validation to balance data fit and physical consistency. Mean squared error loss (data fitting term): In the formula, This represents the number of training samples; For the first i The actual indoor temperature of each sample; This represents the network's predicted value.

[0032] Physical constraint losses (ensuring compliance with RC thermal dynamics): In the formula, , Calculated by self-differentiation.

[0033] The training objective of PINN is to minimize the total loss function and learn network parameters that simultaneously satisfy both statistical data patterns and physical constraints. The fourth step is to evaluate the importance of weights / channels, remove redundant components, retain key structures, and build a compact computing network.

[0034] The fifth step is parameter quantization, which converts high-precision floating-point parameters into low-precision integers to reduce storage and computational overhead.

[0035] The sixth step is to determine the types of network components by reducing the number of parameters, thus balancing performance and efficiency.

[0036] 1) Standard Convolution (Number of Parameters and Computational Cost) Given the input feature map: C in × H in × W in (Number of channels × Height × Width); Convolution kernel: C out × Cin × k × k (Output channels × Input channels × Core size × Core size); Output feature map: C out × H out × W out .

[0037] Parameter count: Param std = C out × C in × k × k Computational cost (FLOPs): FLOP sstd = C out × C in × k × k × H out × W out (2) Depthwise separable convolution (two steps: depthwise convolution + pointwise convolution) Depthwise Convolution: One convolution kernel per input channel; parameter count / computation cost: Param dw = C in ×1× k × k FLOP sdw = C in × k × k × H out × W out Pointwise Conv: 1×1 convolution with adjusted channel count; parameter count / computational cost: Param pw = C out × C in ×1×1, FLOP spw = C out × Cin ×1×1× H out × W out Total number of parameters: Param ds =Param dw +Param pw = C in k 2 + C in C out Total computational cost: FLOP sds =FLOPs dw +FLOP spw = C in k 2 H out W out + C in C out H out W out 3) Efficiency Advantage Comparison Parameter reduction percentage: For example, when k =3 is, Ration param 1 / 9, the number of parameters is reduced to 1 / 9.

[0038] The reduction in computational load is proportional to the reduction in the number of parameters: Ratio flops =Ratio param ≈1 / k 2 .

[0039] Step S3: Based on the baseline value and dynamic trend of air conditioning load in the future period, calculate the adjustable power range, adjustment rate and duration to provide quantitative basis for power grid dispatch.

[0040] Step S4: Continuously monitor the prediction error of the lightweight dynamic hybrid model. When the error exceeds the preset threshold or reaches the set update cycle, trigger the online learning process of the model to adapt to the data distribution drift.

[0041] The process of triggering the online learning process of the model to adapt to data distribution drift includes: Obtain the random variables of the load forecast output from the model, and represent the probability distribution of the model output using a Gaussian distribution; Confidence intervals are calculated based on the probability distribution output by the model. Based on the confidence interval half-width and the reserved additional reserve capacity, obtain the actual executable effective reduction instructions; The sliding window of the dataset used to train the model is dynamically updated based on a preset sliding time window, and the retraining of the lightweight dynamic hybrid model is triggered after the preset time period ends or the number of samples reaches a preset threshold; during full retraining, the model parameters are updated by minimizing the loss of the window dataset.

[0042] In retraining, to avoid the high computational cost of full retraining, incremental learning only uses the newly added sample set to update parameters and introduces a forgetting suppression regularization term.

[0043] Specifically, the online learning process in this embodiment includes: The first step is to introduce a probabilistic layer or Bayesian neural network to output the most likely load prediction value and generate a distribution around that value, which visually shows the range of uncertainty in the prediction, thus transforming the uncertainty from implicit to explicit.

[0044] 1) First, let Y be the random variable for load forecasting (representing the actual load value at a future time period). The probability distribution output by the model is used to characterize the likelihood of Y taking different values. Taking the most commonly used Gaussian distribution (normal distribution) as an example, its probability density function (PDF) is: in, The mean of the distribution corresponds to the most likely load forecast value output by the model; σ is the variance of the distribution (σ is the standard deviation), which quantifies the dispersion of the predicted value (the magnitude of uncertainty); y is the specific value of the load (a continuous variable); p(Y=y) is the probability density of the actual load being equal to y (reflecting the likelihood of that value).

[0045] The core output of a Bayesian neural network (BNN) or probabilistic layer is the parameters (μ, σ) of the aforementioned distribution. 2 Instead of a single, definite value, it explicitly characterizes the uncertainty of prediction.

[0046] 2) Then obtain the confidence interval.

[0047] A confidence interval is a concise quantitative form of a probability distribution, used to give the confidence level that the actual load falls within this range. For a confidence level of 1-α (e.g., α=0.05 corresponds to a 95% confidence level), the confidence interval [L,U] for load Y satisfies: P(L≤Y≤U)=1-α The confidence interval for a Gaussian distribution is calculated as follows: like Then standardized variables (Standard normal distribution). Let z α / 2 The upper α / 2 quantile of the standard normal distribution (e.g., z at a 95% confidence level). 0.025 =1.96), then the confidence interval is: [μ-z α / 2 σ,μ+z α / 2 σ].

[0048] 3) Determine the relationship between the confidence interval width and uncertainty.

[0049] The half-width (range of fluctuation) of the confidence interval is: Δ = (UL) / 2 = z α / 2 σ Δ directly reflects the magnitude of prediction uncertainty; the larger the Δ, the higher the prediction uncertainty.

[0050] 4) Quantitative application of uncertainty in power grid dispatching Let C be the predictive random variable for load reduction potential (representing the amount of load reduction that can be achieved by air conditioning), and its 1-α confidence interval be [C]. L C U ], half-width is ΔC = (C U -C L ) / 2.

[0051] In the scheduling strategy, to address the risk of the reduction effect deviating from expectations, the additional reserve capacity R that needs to be reserved is positively correlated with the half-width of the confidence interval, that is: Wherein, k is the risk coefficient (k≥1, set by the dispatching department according to safety requirements; the higher the uncertainty, the larger the value of k). Δ C The larger the value (the wider the confidence interval), the larger the value of R, ensuring the robustness of the scheduling strategy.

[0052] Furthermore, the practically executable effective reduction index C cmd Must meet: C cmd ≤ C L + R This ensures that even if the reduction potential is at the lower limit of the confidence interval, the scheduling needs can still be met by combining the reserve capacity, thus avoiding insufficient execution effect.

[0053] The second step is to introduce a model retraining mechanism based on a sliding time window, supplemented by an incremental learning strategy.

[0054] 1) Definition of sliding window time and dataset.

[0055] Let the sliding time window length be T (e.g., "the past 30 days", i.e., T=30 days), the current time be t, and the time range covered by the window be [tT, t]. The training dataset within the window is defined as: Where d τ This represents a sample at time τ (containing input feature x). τ With load label y τ , that is, d τ =(x τ ,y τ )).

[0056] The dynamic update rule for the sliding window is: when the data d at time t+1 is updated... t When entering with +1, the window automatically discards the oldest data. t -T, and incorporate new data d t +1, the updated window dataset is: When new data arrives, the oldest data in the window is discarded and the latest data is included to ensure that the dataset always reflects the operation status of the latest time period T.

[0057] 2) Triggering and execution of sliding window retraining First, determine the retraining trigger conditions. Set the retraining trigger period to Δt (or the trigger threshold to k, i.e., trigger once every k new samples). Retraining will start when the following conditions are met: in, For the newly added sample set, || represents the number of samples.

[0058] During full retraining, the model parameters θ are minimized using a window of the dataset. Update the prediction loss (e.g., mean squared error, MSE) on the above in, This represents the load prediction value of the model at time τ under parameter θ.

[0059] 3) Parameter update based on incremental learning strategy with low computational overhead To avoid the high computational cost of full retraining, incremental learning only utilizes the newly added sample set. Update the parameters and introduce a forgetting suppression regularization term (such as Elastic Weighted Consolidation, EWC) to ensure that learned knowledge is not forgotten. The parameter update formula is: The incremental loss term is as follows: This allows for rapid updates by calculating losses only for newly added data.

[0060] The forgetting inhibition regularization term can be expressed as: in, For the model parameters before the update (old knowledge); F i For parameter θ i Importance weight (calculated from the second derivative of the loss function of the old task, measuring the contribution of this parameter to old knowledge); λ>0 is the regularization coefficient, balancing the learning of new knowledge and the retention of old knowledge.

[0061] This method enables rapid updates to its own parameters using only new data, rather than training from scratch, without forgetting previously learned knowledge.

[0062] 4) The combination mechanism of sliding window retraining and incremental learning Based on computing resources and real-time requirements, the model parameter update strategy is as follows: The third step is model self-learning based on meta-learning. The core of meta-learning is learning from "tasks". First, we define the meta-task and meta-dataset for the air conditioning load prediction scenario: 1) Meta-learning tasks and meta-datasets The definition of a meta-task is that each meta-task T corresponds to a load forecasting task for a building type (office building / shopping mall / hotel), and consists of a support set (a small number of samples for rapid adaptation) and a query set (validation samples for evaluating generalization ability): Among them, support set : A small number of load samples corresponding to a certain type of building (K is the number of small samples, such as K=5-20). Input features (temperature / time / building parameters). For load labels; query set : A validation sample corresponding to the building, used to evaluate the task adaptation effect.

[0063] The training dataset consists of multiple meta-tasks (covering various building types) sampled from the task distribution p(T), i.e.: Where N is the number of meta-tasks, and p(T) is the distribution of building type tasks (ensuring coverage of diverse building scenarios).

[0064] 2) Meta-training process The goal of meta-training is not to directly learn load prediction for a specific type of building, but rather to train the meta-model parameters θ to enable it to "quickly adapt to new tasks from a small number of samples." The loss function of meta-training is the average generalization loss over multiple tasks: in, For the meta-model θ via support set S T Task adaptation parameters after a small number of gradient updates (fast fine-tuning, such as 1-5 steps) (α is the fine-tuning learning rate); L() is the load prediction loss (such as mean squared error). ); To meet the expectations of task distribution, the meta-model is designed to ensure general adaptability across diverse tasks.

[0065] 3) Rapid adaptation to new architectural scenarios based on meta-testing When encountering new building scenarios (such as new mission T) new (For building renovation / new building types), only a small number of support sets S are needed for this scenario. Tnew For the meta-model θ Perform rapid fine-tuning to obtain a prediction model adapted to the new scenario: The new task was quickly fine-tuned. in, A small sample of the new building.

[0066] Load forecasting for new scenarios uses a finely tuned model The model predicts the load on new buildings: in, For the prediction function of the meta-model (such as a neural network); Input features for new buildings; To predict load values.

[0067] 4) Cross-scene adaptation model based on transfer learning and meta-learning The combination of meta-learning and transfer learning enables rapid adaptation across architectural scenarios. The overall predictive performance of the framework can be expressed as: Among them Acc new To improve the prediction accuracy of new building scenarios, ensure that the meta-model maintains high prediction reliability even after fine-tuning.

[0068] Furthermore, this embodiment proposes an efficient "cold start" strategy, utilizing prior knowledge of similar buildings to learn and effectively overcome the data bottleneck in the cold start phase, thus providing feasibility for large-scale, rapid deployment of air conditioning load forecasting systems. Details are as follows: Step 1: Cold start parameter initialization strategy based on similar building transfer learning.

[0069] 1) Definition of source / target domain and model parameters.

[0070] Source domain (similar buildings with sufficient data): dataset (N) s (for a large number of samples), source model M s The parameters are divided into general feature layers (learning common patterns across buildings, such as shallow convolutional / recurrent layers) and source building-specific layers (learning personalized patterns of the source building, such as the top fully connected layer): Target domain (new buildings / renovated buildings, data scarce): Dataset ( ), target model M t The parameter structure is consistent with the source model: .

[0071] 2) Initialization of target model parameter migration.

[0072] The general layer directly reuses the common knowledge of the source model (already learned meteorological response and physical laws), while the specific layer is re-initialized: 3) Fine-tuning of the target model (using only a small amount of target data).

[0073] With a fixed general layer (or minor updates), optimize only the specific layer to minimize the prediction loss: This allows for fine-tuning of only the top fully connected layer while keeping the general layer largely unchanged, thus reducing dependence on the target data.

[0074] The second step involves a hybrid model strategy that initially prioritizes physics and gradually augments data.

[0075] 1) Output definition of the physics-data driven hybrid model.

[0076] The hybrid model output is a weighted sum of the physics sub-model and the data-driven sub-model, with the weights dynamically adjusted as the cumulative data volume increases. constraint: (Weighted sum is 1) in: Output for physical sub-models (such as RC thermal models, based on energy conservation:) (output load forecast value) The output of the data-driven sub-model (such as neural network predictions); N(t) is the cumulative amount of target building operation data at time t; The weights are data-driven (increasing with N(t)). The weights of the physical model (decreasing as N(t) increases).

[0077] 2) Dynamic weighting function (evolves with the amount of data).

[0078] Set data volume thresholds: N0 (cold start threshold, such as several days / weeks of data) and N1 (sufficient data threshold), with weights using a saturated increasing function: in, The initial data drives the weights, while β retains the weights for the physical model in the stable phase to ensure physical rationality.

[0079] 3) Optimization objectives (loss functions) at different stages.

[0080] Cold start phase ( Forced data-driven learning of physical laws, the loss function includes physical constraint terms: ( (This penalizes deviations between data-driven and physical models, avoiding serious errors from purely data-driven approaches.) Transition phase ( Gradually reduce physical constraints, focusing on overall accuracy: The coefficient of the physical constraint term decreases linearly with the increase of data volume.

[0081] Stable phase ( Focusing solely on prediction accuracy, the physical model ensures its rationality through its own structure. .

[0082] Example 2 This embodiment, based on Embodiment 1 above, discloses a lightweight dynamic air conditioning load control capability prediction system with embedded physical information. The system is as follows: Figure 2 As shown, it includes: a data input module, a preprocessing and fusion module, a hybrid model prediction module, an adaptive update module, and an output module.

[0083] The data input module is responsible for collecting multi-source heterogeneous data required for forecasting, mainly including outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting building thermal characteristics.

[0084] The preprocessing and fusion module cleans the collected data to remove noise and outliers, and processes missing values ​​through interpolation. For data from different sources and at different frequencies, it performs time alignment and feature extraction to form a unified, high-quality feature vector.

[0085] The hybrid model prediction module is an organically integrated lightweight dynamic hybrid architecture that outputs the basic value and dynamic trend of air conditioning load for future periods based on physical information constraints.

[0086] The adaptive update module continuously monitors the model's prediction error. When the error exceeds a preset threshold or reaches the set update cycle, it triggers the model's online learning process to ensure that the model can adapt to data distribution drift caused by environmental changes, equipment aging, or changes in user behavior patterns, and maintain its long-term prediction accuracy.

[0087] The output module calculates the adjustable power based on the predicted load dynamics.

[0088] For details regarding the above modules, please refer to the relevant descriptions and effects in Example 1 for further understanding.

[0089] Example 3 Based on Embodiment 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the lightweight dynamic air conditioning load control capability prediction method as described above with physical information embedding.

[0090] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the aforementioned lightweight dynamic air conditioning load control capability prediction method with embedded physical information. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0091] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0092] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding, characterized in that, The method includes: Collect and preprocess multi-source heterogeneous data required for prediction. The multi-source heterogeneous data includes outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting the thermal characteristics of buildings. Using the preprocessed multi-source heterogeneous data, and based on physical information constraints, a pre-constructed lightweight dynamic hybrid model is used to output the baseline value and dynamic trend of air conditioning load for future periods. Based on the baseline value and dynamic trend of air conditioning load in the future time period, the adjustable power range, adjustment rate and duration are calculated to provide a quantitative basis for power grid dispatch. The prediction error of the lightweight dynamic hybrid model is continuously monitored. When the error exceeds the preset threshold or reaches the set update cycle, the online learning process of the model is triggered to adapt to the data distribution drift.

2. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 1, characterized in that, The preprocessing process includes: The collected data is cleaned to remove noise and outliers; Missing values ​​are handled using interpolation methods; For data from different sources and at different frequencies, time alignment and feature extraction are performed to form a unified, high-quality feature vector.

3. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 1, characterized in that, The construction process of the lightweight dynamic hybrid model includes: A simplified first-order RC network model is established to describe the dynamic temperature changes in a single area of ​​a building. The first-order RC network model is solved to obtain the physical feature set. A physical neural network is constructed based on the physical feature set, and the physical feature set is used as part of the input features; A lightweight neural network model is established to characterize the dynamic response of air conditioning load. The output of the lightweight neural network model is the predicted indoor temperature. The RC thermal equilibrium differential equation is embedded as a physical constraint term into the loss function of the lightweight neural network model. The loss function is a weighted sum of data loss and physical constraint loss. With the goal of minimizing the loss function, the lightweight neural network model is trained using the loss function, and finally a pre-built lightweight dynamic hybrid model is obtained.

4. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 3, characterized in that, After training the lightweight neural network model, redundant components are removed and key structures are retained through weight / channel importance evaluation to construct a compact computational network; the high-precision floating-point parameters of the lightweight neural network model are converted into low-precision integers; and the types of network components are determined by the parameter reduction ratio.

5. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 3, characterized in that, The expression for the loss function is as follows: in, These are the physical constraint weighting coefficients. The number of training samples, For the first i The actual indoor temperature of each sample For the network prediction value, The equivalent heat capacity of the building's interior space. The equivalent thermal resistance of the building envelope. and Indoor and outdoor temperatures, respectively. The cooling or heating power provided to the air conditioning system.

6. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 3, characterized in that, The expression for the first-order RC network model is: in, The equivalent heat capacity of the building's interior space. The equivalent thermal resistance of the building envelope. and Indoor and outdoor temperatures, respectively. The cooling or heating power provided to the air conditioning system, Heat is gained internally.

7. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 1, characterized in that, The process of triggering the online learning model to adapt to data distribution drift includes: Obtain the random variables of the load forecast output from the model, and represent the probability distribution of the model output using a Gaussian distribution; Based on the probability distribution output by the model, the confidence interval is calculated and obtained; Based on the confidence interval half-width and the reserved additional reserve capacity, obtain the actual executable effective reduction instructions; The sliding window of the dataset used to train the model is dynamically updated based on a preset sliding time window, and the lightweight dynamic hybrid model is retrained after the preset time period ends or the number of samples reaches a preset threshold; during full retraining, the model parameters are updated by minimizing the loss of the window dataset.

8. The lightweight dynamic air conditioning load regulation capability prediction method based on physical information embedding according to claim 7, characterized in that, In the retraining process, to avoid the high computational cost of full retraining, incremental learning only uses the newly added sample set to update parameters and introduces a forgetting suppression regularization term.

9. A lightweight dynamic air conditioning load regulation capability prediction system with embedded physical information, characterized in that, The system includes: a data input module, a preprocessing and fusion module, a hybrid model prediction module, an adaptive update module, and an output module; The data input module is responsible for collecting multi-source heterogeneous data required for prediction, mainly including outdoor meteorological data, historical and real-time load data of building air conditioning systems, and static parameters reflecting building thermal characteristics. The preprocessing and fusion module cleans the collected data to remove noise and outliers, and processes missing values ​​through interpolation. For data from different sources and at different frequencies, it performs time alignment and feature extraction to form a unified, high-quality feature vector. The hybrid model prediction module is an organically integrated lightweight dynamic hybrid architecture that outputs the basic value and dynamic trend of air conditioning load for future periods based on physical information constraints. The adaptive update module continuously monitors the model's prediction error. When the error exceeds a preset threshold or reaches the set update cycle, it triggers the model's online learning process to ensure that the model can adapt to data distribution drift caused by environmental changes, equipment aging, or changes in user behavior patterns, and maintain its long-term prediction accuracy. The output module calculates the adjustable power based on the predicted load dynamics.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the lightweight dynamic air conditioning load control capability prediction method with physical information embedding as described in any one of claims 1-8.

Images