Air conditioner load large model prediction method considering external environment and irrational behaviors

Abstract

The invention relates to the technical field of power systems, in particular to an air conditioner load large model prediction method considering an external environment and irrational behaviors, comprising the following steps: S1, acquiring resident attributes and text information, and constructing static and semantic portrait vectors; s2, collecting environment data and generating a semantic scene prompt; s3, inputting into a large language model to generate an air conditioner control script; s4, converting the script into a power sequence, and fusing external and behavior factors; s5, correcting the power curve based on thermal inertia and equipment constraint; s6, constructing a virtual agent and superposing the virtual agent to generate a group load; s7, carrying out Monte Carlo modeling and carrying out online correction on prediction deviation; and S8, outputting a mean value, a risk and a response suggestion for scheduling. According to the method, the user behavior script is generated by introducing the large language model, and the multi-layer translation mechanism is constructed, so that high-precision prediction and adjustment suggestion output of the group air conditioner load are realized, and the prediction responsiveness, interpretability and scheduling practicability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a large-scale model prediction method for air conditioning load that takes into account external environment and irrational behavior. Background Technology

[0002] Electricity load forecasting is a core technical link in ensuring the safe, stable and economical operation of the power system. Especially under the dual carbon targets, the electrification trend of residential loads is obvious. Air conditioning loads, which are temperature-controlled, have become the main factor causing the peak-valley difference of the power grid during the high-temperature period in cities. Due to their significant proportion of peak load, accurate air conditioning load forecasting is of great significance for the grid's peak shaving and valley filling, demand response and virtual power plant scheduling decisions. In recent years, with the development of artificial intelligence technology, deep learning-based electricity load forecasting methods have gradually replaced traditional linear models and become the mainstream of research and application, driving the continuous evolution of forecasting models in time series modeling and multi-source data fusion.

[0003] Most existing load forecasting methods rely on time series modeling or optimization decision frameworks based on utility functions. While the former can capture historical patterns, it lacks the ability to respond to sudden behavioral changes such as holiday travel, extreme heat, or load fluctuations caused by social events, often resulting in systematic biases in peak forecasting. The latter, based on rational economic assumptions, cannot cover the behavioral characteristics of residents driven by irrational factors such as emotions and herd mentality in the real environment, leading to structural errors in forecasting. In addition, most models rely on manually defined feature construction and parameter tuning, resulting in lag in updates and poor adaptability when facing complex social situations and rapidly changing external environments, making it difficult to support the scheduling system's forecasting requirements for high timeliness, interpretability, and flexibility. Summary of the Invention

[0004] This invention provides a large-scale model prediction method for air conditioning load that considers external environment and irrational behavior. Based on a large language model, it takes resident profiles and environmental contexts as inputs to directly generate air conditioning control scripts that conform to life logic. It breaks away from the reliance of traditional models on rational assumptions. Through semantic enhancement and retrieval mechanisms, it integrates multi-source heterogeneous information such as meteorological data, electricity prices, and social corpora to improve the model's environmental adaptability and behavior reproduction ability. Combined with physical constraints and thermal inertia models, it ensures the executability and physical consistency of the prediction results. At the same time, it introduces online calibration and uncertainty assessment mechanisms to improve the model's stability and prediction accuracy under abnormal scenarios, achieving highly reliable air conditioning load prediction that meets the heterogeneity of resident behavior and the needs of the scheduling system.

[0005] A large-scale model prediction method for air conditioning load considering external environment and irrational behavior includes the following steps: S1: Collect residents' basic information and generate static profile vectors. If there is historical natural language text data, extract semantic vectors. After fusion, form the final individual profile vector. After logical verification, large language model completion and clustering, output numerical profile vectors and semantic profile summary text. S2, collect environmental data within the prediction time window, construct environmental vector sequences and scenario prompt text, and extract behavioral fragments similar to the current scenario; S3, the generated numerical profile vector, semantic profile summary text, temporal environment vector and contextual prompt text are input into the large language model to generate an air conditioning control script including timestamp, operation type and parameters; S4 parses the air conditioning control script into structured operation tuples, combines the external temperature, set temperature, wind speed and mode, calculates the instantaneous power of the air conditioner, and integrates the irrational coefficient and social influence to form an initial power sequence; S5, based on thermal inertia and equipment limitations, performs physical consistency correction on the initial power curve, constrains the start-stop frequency and maximum power fluctuation, and forms a corrected power curve; S6: Sample resident profiles to generate virtual resident agents, overlay power output and introduce social influence to generate predicted values ​​of group air conditioning load; S7 uses Monte Carlo sampling to construct an uncertainty prediction set and performs online learning and short-time error correction when prediction bias occurs; S8 outputs the mean, standard deviation, and confidence interval of the predicted air conditioning load for the group, generating risk assessment and adjustment recommendations.

[0006] Optionally, S1 includes: S11, collect basic information for each resident, including age. ,income ,Profession Housing type Family structure and historical energy consumption statistics Generate resident attribute vectors , is represented as: ; S12, the resident attribute vector is input into the feature embedding network for encoding, generating a static profile vector. , is represented as: ; in, Encoding function for static images; S13, If residents have historical natural language text data (such as customer service dialogues, service records, etc.), use a semantic encoder to extract their semantic vectors. , is represented as: ; in, For user-related text collections, For semantic text encoding functions; S14 concatenates the static portrait vector with the semantic vector and maps them to a unified semantic feature space to form the final individual portrait vector. , is represented as: ; S15, the final individual profile vector undergoes logical relationship verification (e.g., whether age and occupation are reasonably matched) and statistical anomaly detection. Outlier detection is performed on the feature dimensions using standard deviation normalization, as shown below: ; when When this occurs, a correction or marking as an abnormal user is triggered; S16, for users with missing or abnormal text, a large language model is invoked to generate complete text, and clustering matching is performed based on the embedding results and historical image vectors in the image database. Cosine similarity is used to define the semantic similarity of the images, expressed as: ; when If the image deviates significantly from the known group, it indicates that manual verification or automatic re-inference is required. The similarity threshold; S17, Output numerical portrait vector (final individual portrait vector) ) and semantic summary text.

[0007] Optionally, S2 includes: S21, Set the prediction time window Extracting external temperature data from multi-source databases relative humidity Wind speed Solar radiation intensity Air Quality Index Time-of-use electricity pricing Demand response signals Event (the event database provides a sequence of tags for holidays, sporting events, exams, performances, etc.) And the frequency of social keywords (social databases are crawled and clustered with topics to extract keywords such as "air conditioner", "heat wave", and "cooling off") Environmental feature vectors are constructed from the environmental data, including those from the environment. , is represented as: ; S22, for environmental feature vectors Perform normalization and smoothing operations for each feature dimension. Using sliding window mean filtering and normalization transformation, it can be expressed as: ; ; in, To smooth the window width, , These are the historical mean and standard deviation; S23, All normalized environmental feature vectors are combined into an environmental vector sequence, represented as: ; S24, Use the semantic mapping module to convert the environment vector sequence into semantic contextual cue text. ; S25 introduces a Retrieval Enhancement Generation (RAG) mechanism based on a large language model, which calculates query vectors through a semantic indexing model. It then retrieves historical behavior segments similar to the current environmental context from the constructed environmental corpus, calculates similarity using vector cosine similarity, and selects the segments with the highest similarity. The results of a similar scenario search are represented as follows: ; S26, Output the numerical environment vector sequence and semantic contextual cue text .

[0008] Optionally, S3 includes: S31, for numerical image vectors Semantic summary text Numericalized environment vector sequence and semantic contextual cue text This concatenation is used to drive the Large Language Model (LLM) to generate event sequences. Each event consists of a triple, represented as: ; in, For timestamps, As an action type, For parameters (such as set temperature, duration, wind speed, mode, etc.); S32, the script event sequence is generated through maximum likelihood estimation, represented as: ; in, These are learnable parameters; S33, supervised training is performed by minimizing the standard cross-entropy using instruction fine-tuning (SFT), as follows: ; S34 introduces an irrational behavior mapping function to define the irrationality coefficient of an individual's response to the environment, expressed as: ; in, This is a statistical summary of the environment vectors within the time window. As an indicator of sentiment or conformity in the intensity of social trending words, For the Sigmoid function; The temperature parameter is set as follows: ; S35 introduces feasibility and vocabulary constraints, defining a feasibility penalty function as follows: ; in, A large penalty coefficient, For indicator functions; S36, apply vocabulary sparsity regularization to the generated parameters and introduce a temperature coverage regularization term to control and adjust the sparsity, expressed as: ; ; in, For a set of sliding time windows, The minimum number of events threshold; S37, through joint optimization of behavior generation loss, temperature coverage regularization, and physical consistency loss, is expressed as: ; ; S38 defines a timestamp position error function to control the accuracy of event generation and outputs the script sequence and its corresponding execution confidence, expressed as: .

[0009] Optionally, S4 includes: S41, perform semantic parsing and structured processing on the natural language segments in the air conditioning control script, generating a sequence including timestamps. Operation type Parameter values and operating mode The Quadruple Event , is represented as: ; ; in, This is a mapping function from language embeddings to parameter differences; S42, based on the operation type, set parameters, and environmental conditions, calculates the instantaneous power of the air conditioner, expressed as: ; in, This refers to standby power. This is the power-on indicator function. These are the equipment's characteristic parameters; S43, using the pattern mapping function The modified energy efficiency parameter set is expressed as follows: ; S44 introduces a thermal balance differential equation to model the dynamic response of air conditioning operation to indoor temperature, expressed as: ; ; S45, introducing an irrational correction coefficient. And the power is corrected once, expressed as: ; S46, introduces a sliding fusion mechanism to compare power with the average power of the local population. The power is fused to form a smoothed power estimate, expressed as: ; in, The social coupling coefficient; S47, perform a sliding integral on the power sequence to obtain the final smoothed power output, and apply boundary constraints, as follows: ; ; S48 outputs the final time series power data, represented as follows: .

[0010] Optionally, S5 includes: S51 introduces a simplified first-order heat balance equation to simulate the change process of indoor temperature, expressed as: ; in, For the heat capacity of the building, For thermal resistance, This refers to the cooling efficiency factor of the air conditioner. This refers to the instantaneous power of the air conditioner. S52 introduces a physical consistency correction term to correct the power response where the current power and indoor temperature deviation directions are inconsistent, expressed as: ; in, This is the power-on indicator. To prevent small constants from being divided by zero; S53 sets the maximum and minimum operating power for the air conditioning equipment, expressed as: ; S54, Set minimum start-stop interval With maximum adjustment range , is represented as: ; S55, the power series is smoothed using a moving average method, expressed as: ; in, To adjust the sliding window size; S56 introduces physical consistency and device constraints to correct the power sequence, correct unreasonable temperature control response behavior, and generate a corrected power curve.

[0011] Optionally, S6 includes: S61 samples profile information from different users, generates virtual resident agents using a Large Language Model (LLM), and outputs the air conditioning load of the virtual agents. ; S62, perform a weighted summation of the air conditioning loads of all virtual agents to form a group load estimate. , is represented as: ; in, as an agent The weight, The number of agents in the group; S63 introduces a social keyword frequency correction factor to construct social coupling. , is represented as: ; in, It is the Sigmoid activation function. This is the adjustment coefficient; S64 introduces a correction factor for group load. The final predicted population load is constructed by incorporating a social coupling correction factor, and is expressed as: .

[0012] Optionally, S7 includes: S71 will extract user profiles and environmental data, construct a perturbation distribution script, perform Monte Carlo sampling from it, and generate multiple sets of sample power curves. And calculate the population average predicted power. , is represented as: ; in, The number of samples generated; S72, at each actual sampling time Calculate the predicted value Compared with the actual measured value The difference , is represented as: ; Based on difference Define the calibration loss function as follows: ; S73 introduces Kalman filtering to dynamically adjust the predicted power, expressed as: ; S74, introduces a regularization term. To avoid overfitting, the final calibration loss function is expressed as: ; in, This is the regularization coefficient.

[0013] Optionally, S8 includes: S81, in each time period The above generates predicted air conditioning load values ​​from multiple samples. The average value is calculated to reflect the overall load trend at that moment, and is expressed as: ; in, The number of samples generated; S82, calculate the standard deviation of the air conditioning load forecast results for each time period, and construct the confidence interval, as follows: ; ; ; in, For the first Forecasted average air conditioning load for the time period; S83, taking into account the sensitivity to temperature and load changes, generates load regulation recommendations, expressed as follows: ; in, For load changes, For temperature adjustment range, Power sensitivity related to temperature changes, This is the base power of the air conditioner; S84 calculates load risk indicators using the standard deviation and average value of air conditioning load forecasts. , is represented as: ; S85, based on the predicted air conditioning load and time-of-use electricity pricing, calculates the cost-effectiveness for each time period, expressed as: ; S86 ultimately outputs the predicted air conditioning load, standard deviation, confidence interval, load risk index, cost-effectiveness, and adjustment recommendations.

[0014] The beneficial effects of this invention are: This invention introduces a Large Language Model (LLM) to perform semantic understanding and reasoning modeling of user profiles and the external environment. Based on natural language, it can generate air conditioning usage scripts that conform to user behavior logic, realistically depicting residents' irrational temperature adjustment behaviors in situations such as extreme weather, energy-saving responses, and social conformity. This effectively overcomes the limitations of traditional load forecasting, which portrays user behavior as overly rational and static. Relying on the semantic generation and retrieval enhancement mechanism (RAG) of the Large Language Model, the system can dynamically integrate multi-source unstructured data such as meteorological data, electricity prices, air quality data, and social corpora to form a semantically interpretable prediction context, thereby improving the responsiveness and situational adaptability of load forecasting.

[0015] This invention constructs a multi-layered translation mechanism from user behavior scripts to power curves and then to group load forecasting. In the process, it introduces thermophysical models, equipment constraints, social coupling, and uncertainty modeling. This not only generates highly fitted individual power sequences, but also outputs confidence intervals, risk indicators, adjustment suggestions, and economic evaluation results. This meets the refined control needs of power systems in scenarios such as load dispatching, demand response, and peak-valley reduction. At the same time, with the help of online calibration and parameter self-updating mechanisms, the system has continuous learning and self-adaptation capabilities, improving the long-term stability and practicality of the forecasts. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the prediction method according to an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0019] like Figure 1 As shown, the large-scale model prediction method for air conditioning load, considering external environment and irrational behavior, includes the following steps: 1. Profile Construction and Verification: At this stage, the system aims to generate an individual profile vector and a corresponding natural language profile summary for each resident user, which can be understood and reasoned by the large language model. The profile should reflect the individual's static characteristics (age, income, occupation, type of residential building, etc.) and encode their behavioral preferences and irrational tendencies over time, so that it can be used by the LLM in the subsequent behavior generation stage.

[0020] Profile building first receives basic input from resident profiles or power company databases, including resident attribute vectors: in, Indicates age group, Annual income range Indicates occupational category, For housing type and building features (such as orientation and insulation coefficient). Indicates family member structure (including markers for elderly or children). These are historical energy consumption characteristics statistics (such as mean, variance, number of start-stop cycles, etc.).

[0021] Next, the model maps the structured data to a continuous space using a feature embedding network: in, This is a static image encoding function, which can be a multilayer perceptron or a Transformer encoder.

[0022] Meanwhile, if there are historical natural language descriptions of users (such as questionnaires, service records, customer service conversations), the system extracts their text vectors through a semantic encoding model: in, This is a collection of user-related text. The final individual profile vector is obtained by concatenating static and semantic features and then applying a non-linear mapping layer. This vector, as an embedding of residents in the LLM input, reflects their typical behavioral patterns and potential preferences.

[0023] To ensure the reasonableness of the profile, the system performs multi-layered verification after construction. First, physical and statistical constraints are checked on the structural domains, such as whether the combination of age and occupation is statistically consistent (if a resident's occupation is "student," then the age range should be below 30); logical consistency is checked on the matching relationship between income and housing area. Second, regularization correction is performed on energy consumption characteristics. This is done using historical average energy consumption data. with standard deviation Calculate the normalized index: And outliers are detected using confidence intervals. If so, it will trigger a correction or mark the user as an abnormal user.

[0024] Building upon this foundation, to compensate for missing or ambiguous resident attributes, this invention utilizes the semantic completion capabilities of a large language model to perform contextual inference. Input templates such as: "Resident is of {occupation}, located in {region}, income information not provided, please complete a reasonable income range based on regional average and occupation type," are used by the LLM to generate the completed description. The completed result is then mapped back to vector space via a numerical analysis module, ensuring the profile is complete.

[0025] Finally, the system uses the defined similarity metric: Clustering verification is performed on the current image and historical image sets. If the similarity with any existing class center is below a threshold... If the result is negative, it indicates that the portrait differs significantly from the known group and requires manual verification or automatic re-inference.

[0026] After the above steps, the profile construction and verification stage outputs two types of results: one is semantic profile summary text for LLM use, and the other is numerical profile vector. Together, these two elements form the foundational input for generating subsequent behavior scripts, enabling the model to fully reflect individual characteristics and potential irrational preferences when generating residents' air conditioning usage behavior, while ensuring data integrity, logical consistency, and statistical reasonableness.

[0027] 2. Environmental Context Retrieval: After the profile is constructed, the system enters the environmental context retrieval stage. Its goal is to construct external environmental context vectors and semantic descriptions for specific residents or resident groups within the prediction period, enabling the large language model to generate air conditioning usage behaviors that conform to real-world contexts. The core of this stage lies in integrating multi-source, heterogeneous, and asynchronously updated external environmental data into a unified time series and textual contextual prompts, thereby establishing a "human-environment" correspondence.

[0028] Environmental scenario retrieval first determines the prediction time window. The system extracts information from multiple databases, including meteorological, market, event, and social data. The meteorological database provides hourly external temperature data. relative humidity Wind speed Solar radiation intensity and air quality index The market database includes time-of-use electricity prices. With demand response signals The event database provides tag sequences for holidays, sporting events, exams, performances, etc. The social database extracts the frequency of keywords such as "air conditioner," "heat wave," and "cooling off" through web scraping and topic clustering. All data is timestamped and interpolated to a fixed step size. .

[0029] At the numerical level, the system constructs environmental feature vectors: Then, the time-series feature is normalized and smoothed. For each feature... Sliding window mean filtering and normalization transformation are used: in, To smooth the window width, Here, represents the historical mean and standard deviation. The normalized environmental characteristic sequence is denoted as: To enable LLM to understand this multidimensional data, the system introduces a scenario generation submodule. This module uses templated semantic mapping functions. Transform numerical features into natural language descriptions. For example, when Continuously exceeding 35°C and When the value is above 150, the following description is automatically generated: "The current area is experiencing sustained high temperatures and poor air quality; residents tend to keep their windows closed and turn on the air conditioning for extended periods." It peaked between 17:00 and 20:00. When a DR load reduction signal is issued, the following description is generated: "Power rationing is in effect during the evening peak electricity pricing period. Some residents may delay turning on their appliances or raise their set temperatures." These text fragments are combined to form a complete scenario-based warning text. .

[0030] To balance real-time performance and contextual coherence during contextual retrieval, the system employs a Retrieval-Augmented Generation (RAG) mechanism. First, a query vector is calculated using a semantic indexing model. : Then, similar contexts are searched in the environmental corpus, and their relevance is defined by cosine similarity: Take the first one with the highest similarity The retrieved historical scenarios and behavioral fragments are summarized into "typical behavioral responses" using LLM and then concatenated with current scenario features to enhance the model's ability to generate rare or extreme cases.

[0031] After the scenario retrieval is completed, the system performs consistency and logic checks. These checks include time continuity checks, ensuring that changes in meteorological characteristics over a short period conform to physical laws (e.g., the rate of temperature change does not exceed a threshold). The system also performs semantic consistency checks to prevent conflicts between "holidays" and "working days" labels. When a conflict is detected, the system triggers a correction rule: relabeling the data based on the government's holiday announcement database. .

[0032] Finally, the environmental scenario retrieval stage outputs two types of information: one is a numerical environmental vector sequence. The first is for subsequent power translation and correction modules; the second is semantic contextual prompt text. This serves as the contextual input for the behavior generation stage of the large language model. This stage ensures that the model can generate air conditioning usage behavior predictions that are highly consistent with real-world situations, based on an understanding of specific resident profiles and combined with current meteorological, market, and social environmental conditions.

[0033] 3. Behavior script generation (LLM instruction fine-tuning): This stage uses "profile summary text + profile vector embedding" and "contextual prompt text + environmental vector sequence" as joint conditions to drive a large language model to generate timestamped air conditioning usage scripts covering the prediction time window. The generation goal is to directly produce residents' operation sequences rather than solving any explicit utility optimization problem. Irrational behaviors are explicitly manifested in the generation process through data-driven language patterns and control signals.

[0034] The model first constructs the conditional input. The image side provides a semantic summary. With Embedded The environment provides the semantic context. with sequence Combine the two to create a command-style prompt. And through the adaptation layer and The key-value states (such as prefix vectors or retrieval fragments) injected into the model yield conditional representations. Behavioral scripts are represented as sequences of events. Each event Includes timestamps, action type, and parameters (such as set temperature, duration, wind speed, and mode). The generation process uses autoregressive decomposition. in, These are learnable parameters. During the training phase, instruction fine-tuning (SFT) is used to minimize the standard cross-entropy: The labeled alignment data comes from the triplet corpus of "thermostat / energy consumption log → event extraction" and natural language scripts that have been manually or semi-automatically proofread. To ensure that the output can be parsed by the downstream translator, the script adopts a conventional syntax, and the model follows a fixed template during decoding, for example: [07:30 Power-on setting = 26℃ fan speed = 1 mode = cooling] [12:00 shutdown time = 45 minutes] [17:40 Power-on setting = 23℃ duration = 30 minutes subsequent = return to 26℃] [22:30 Sleep Setting = 27℃] Irrational behavior is shaped by both control signals and data distribution, rather than by any utility function. An irrationality coefficient is defined based on the profile and the environment: in, This is a statistical summary of the environment vectors within the time window. For sentiment / conformity indicators such as the intensity of trending social words, It is Sigmoid. During reasoning, it will be... Mapped to decoding temperature and diversity control: This makes script-like patterns such as "rapid cooling," "frequent fine-tuning," and "forgetting to shut down" more likely to occur under extreme high temperatures or strong social conformity. To further enable explicit control, control markers (such as...) are added to the prompts. <irr-high> 、 <price-sensitive-low>These tags are aligned with log facts in natural language descriptions in the training set, and the model learns the mapping of "irrationality intensity → behavioral style".

[0035] To prevent text generation from deviating from physical and device boundaries, a constraint search is used for decoding. Let the action word list be... Equipment and thermal process constraints are indicated by function This means that the constraint search scores the candidates at each step: in, The constraints include minimum start / stop intervals, set temperature step sizes, and mode mutual exclusion, with a large penalty coefficient. Masking is applied to the time and parameter sub-vocabularies. Set the logarithmic probability of the out-of-bounds token to 0. Implement hard constraint masks: To encourage scripts to have a reasonable operation density during high-temperature periods, a coverage regularization is introduced in the generation of long time windows. in, For a set of sliding time windows, The threshold is the minimum number of events. The total loss is jointly optimized using multiple constraints: The script is approximated by rapidly passing through a first-order thermal process and the sequence that clearly violates temperature inertia (such as extremely frequent start-stop) is penalized.

[0036] To ensure the generated scripts are both "human-like" and "executable," preference alignment is employed in the later stages of training. Script pairs are constructed. and reference model Maximizing based on Direct Preference Optimization (DPO): The preference labels are derived by combining heuristic scoring and expert scoring, taking into account physical feasibility, start-stop frequency, comfort threshold, and operational interpretability. This alignment process also does not introduce a utility function; it only learns language patterns that "better align with real human behavior."

[0037] Retrieval enhancements are input along with the instructions during generation. The query vector for the current context. Similar scenes from history Similarity is calculated and concatenated into searchable evidence, enabling the model to reuse similar precedents even in rare events. To ensure temporal consistency, the timeline is set at a fixed granularity. Discrete, the model prioritizes generating relative time offsets. And accumulate to obtain The vocabulary for time is smaller and more stable than that for time, and the accuracy of time positioning is improved during training by using formula xx.

[0038] In addition to the script itself, the inference output also synchronously provides the generated confidence scores. A confidence score is defined for each event. The confidence level of the entire sequence is calculated using the geometric mean. In uncertainty sampling, by... Apply small perturbations and temperature fluctuations, and repeat generation. Scripts, forming a script collection The power distribution obtained subsequently through action-power translation originates from the propagation of this set. The entire process directly derives the resident behavior script from the input profile and scenario. Irrational factors are explicitly injected with linguistic priors and control signals, avoiding any utility maximization assumptions. At the same time, constraint decoding and physical regularization ensure that the script is feasible for the device and facilitates downstream power translation.

[0039] 4. Motion-Power Translation: After the behavior script is generated, the system enters the action-power translation stage, which transforms the residential air conditioning operation script output by the large language model into the corresponding time-of-use power sequence, thereby achieving an interpretable mapping from natural language to physical energy consumption. This stage no longer relies on traditional statistical regression models, but instead achieves multi-layered translation from language to power through semantic parsing, thermophysical calculations, and correction of irrational behaviors.

[0040] The behavior script first undergoes structured processing by the parsing module. The script is typically represented as a language fragment with time and parameter descriptions, such as "turn on the air conditioner at 7:00 AM and set it to 26℃" or "turn off the computer for half an hour during lunch break." The system uses a regularized language parser to convert these into event quadruples: in, Indicates the time when the action occurs. Indicates the operation type (such as power on, power off, temperature adjustment). This represents the set of temperature and wind speed parameters. This refers to the operating mode (cooling, dehumidifying, or ventilation, etc.). When the language expression contains ambiguous elements, such as "turn the temperature down a little," the system utilizes a semantic completion network. Estimate the relative change: in, This is a mapping function from language embeddings to parameter differences, used to ensure that fuzzy operations in the script can be quantified.

[0041] After obtaining structured actions, the system enters the physical characteristic mapping stage. The main factors affecting air conditioning power include on / off state, temperature difference, fan speed, and mode characteristics. The external temperature is defined as... The indoor temperature is set to Wind speed level is The instantaneous power of the air conditioner is given by the following formula: in, This refers to standby power. This is the power-on indicator function. These are device characteristic parameters. If the current event involves mode switching, the system uses a mapping function. Adjust the parameters: To reflect the changes in energy efficiency characteristics under different working modes.

[0042] To characterize the thermal inertia of a dwelling, the system introduces a simplified first-order thermal balance equation: in, For heat capacity, For thermal resistance, Let be the refrigeration efficiency coefficient. Through Euler discretization, the recurrence relation can be obtained: Therefore, the dynamic change trajectory of air conditioning power can be iteratively solved over time series.

[0043] Based on this, in order to reflect the effect of residents' irrational behavior, the system introduces the irrationality coefficient generated by the language model in the previous stage. This is used as a power offset correction term. The specific form is: in, This is an irrational weighting coefficient. When temperatures are high or social emotions are intense, If the power level is too high, the system will automatically increase the power level to simulate "overcooling" or "delayed shutdown".

[0044] At the group level, to describe the social conformity effect among residents, the system establishes a group average power feedback mechanism. Let the average power of the local group be... The single-unit power is then further corrected as follows: in, This is the social coupling coefficient, which dynamically adjusts with the strength of social signals. When the topic of "escaping the heat" is active in public opinion or social media, As the value increases, residents' behavior tends to become more consistent.

[0045] Because language-generated scripts can lead to frequent start-stop cycles and power spikes, the system applies a time-domain smoothing operator to the power sequence to eliminate non-physical abrupt changes. The smoothed power is defined by a moving average: Boundary constraints are applied to the results to ensure that: To comply with equipment operation safety restrictions.

[0046] Finally, the action-power translation module outputs time-series power data: Its time resolution is consistent with the environmental input. The results can be used not only directly as single-user load forecasts but also for group load aggregation and demand response simulations. Through a combination of linguistic parsing, thermophysical calculations, and irrational corrections, this stage achieves a continuous and interpretable transformation of resident behavior from verbal expression to power signals, providing an accurate individual power basis for subsequent group aggregation forecasts.

[0047] 5. Physical consistency and equipment constraint correction: To ensure that the proposed large-scale air conditioning load prediction model, which takes into account external environment and irrational behavior, can operate stably in residential electricity consumption scenarios, this invention designs a complete training and deployment process at the implementation level, including the construction of data processing pipelines, model parameter optimization strategies, online learning mechanisms, and edge-side real-time inference schemes.

[0048] After the action-power translation is completed, the system needs to ensure that the generated air conditioning power curve meets the physical and equipment constraints, avoiding unreasonable power fluctuations or violations of equipment operating limits. The main task at this stage is to correct the power sequence through the physical model and equipment capacity to ensure that the final output power curve meets both the requirements of the thermophysical process and the hard limitations of the equipment, such as minimum start-stop intervals, maximum power, and equipment load capacity.

[0049] First, for each air conditioning operating state, the system needs to ensure that the generated power curve conforms to the thermophysical dynamics of the air conditioning system. Temperature change in air conditioning is a gradual process and cannot be instantaneous. To simulate this process, the system uses a first-order thermal process model to describe indoor temperature changes. The simplified first-order heat balance equation is: in, For the heat capacity of the building, For thermal resistance, This refers to the cooling efficiency factor of the air conditioner. Let be the instantaneous power of the air conditioner. This equation shows a direct relationship between the rate of change of indoor temperature and the outdoor temperature, the current indoor temperature, and the air conditioner power. The system uses numerical integration (such as the Euler method) to solve for the evolution of indoor temperature, and then adjusts the air conditioner power inversely based on the temperature difference. in, This is the power-on indicator. To prevent small constants from being divided by zero, this correction ensures physical consistency between power output and actual indoor temperature variations.

[0050] Secondly, equipment constraint calibration is another crucial step. In actual use, air conditioning equipment has several rigid limitations, such as maximum power limits, minimum start-stop intervals, and equipment load capacity. To prevent exceeding the equipment's operating range, the system first defines the upper and lower power limits for the air conditioner: in, and These represent the minimum and maximum power of the device, respectively. When generating the power curve, if the power exceeds the maximum power or falls below the minimum power at a certain moment, the system will constrain the power to restore it to an acceptable range.

[0051] Furthermore, the start and stop of the air conditioning equipment are not instantaneous operations. To avoid unstable loads caused by frequent start and stop of the air conditioning, the system corrects for each start and stop event. A minimum start and stop interval is set. And monitor power changes: in, This indicates the maximum allowable value for each power adjustment. To minimize the start-stop interval, ensure that power changes do not exceed the set maximum range, and avoid continuous start-stop of equipment within a short period of time.

[0052] If excessive fluctuations occur in the generated power curve within a short period, the system will reduce these unreasonable power fluctuations by adjusting parameters, adding smoothing operations, or merging adjacent temperature adjustment actions. For example, when the system detects frequent temperature adjustment actions or start-stop phenomena in the power curve, it will use a moving average method to smooth the power sequence. in, This is the sliding window size. This method can effectively reduce power spikes caused by irrational behavior (such as frequent temperature adjustments).

[0053] In addition, the system needs to ensure that the air conditioning system does not operate beyond the equipment's rated operating time. For example, the minimum rest time for the air conditioning compressor is usually set to... If continuous operation for a short period of time causes the compressor to be overloaded, the system will forcibly adjust the power to a level lower than the rated power until the equipment can return to normal.

[0054] Finally, to ensure the consistency between the power curve and the actual operating behavior of the equipment, the system provides feedback adjustments to the correction results for physical consistency and equipment constraints, and monitors in real time whether the output power sequence meets the constraints. If any non-compliance is detected, the system will automatically adjust the power sequence to ensure that each generated power curve conforms to both physical laws and equipment operating constraints.

[0055] Through the physical consistency and equipment constraint correction module, the system can ensure the accuracy of air conditioning load forecasting while avoiding the generation of power sequences that do not conform to actual equipment capacity or physical laws. This process ensures the feasibility and stability of air conditioning load, providing a reliable data foundation for subsequent load forecasting and demand response scheduling.

[0056] 6. Group expansion and overlay: The main goal of the group expansion and overlay phase is to extrapolate from the behavior predictions of individual users to the entire resident population, and then estimate the air conditioning load over a larger area. This process not only requires merging the behaviors of multiple individuals into group-level predictions, but also needs to handle the heterogeneity between groups and the collective effects of irrational behaviors to ensure that the estimation of group load is both accurate and feasible.

[0057] First, the system samples user profiles to generate multiple "virtual resident agents," which simulate user behavior characteristics. The profiles and environmental contexts of these virtual agents are used to generate behavioral scripts using a Large Language Model (LLM). Then, the air conditioning operation sequence of each virtual agent is converted into a power curve. Each virtual agent in the group is defined... The load is Therefore, the estimate of the group load is the weighted sum of the loads of all individuals: in, as an agent The weight, It is a group in time Total load, The number of agents in the group. Agent weight. Adjustments can be made based on various factors, such as family income, living space, and number of air conditioners.

[0058] In the group expansion process, the system stratifies the group using profile data and determines the virtual agent generation strategy based on the characteristics of different group levels. User behavior often shows significant differences in air conditioning usage across different household types, income levels, and geographical locations. To make load forecasting more accurate, the system dynamically allocates different virtual agents based on resident characteristic data (such as housing type, region, and seasonal variations). For example, high-income households may prefer to use more air conditioning units and run them for longer periods, while low-income households may use energy-saving modes more frequently.

[0059] Furthermore, the system needs to consider the collective effect of irrational behavior within a group. In large-scale groups, individual irrational behavior is often amplified through social influence or collective reactions triggered by the external environment. For example, during extreme weather events, some residents may be inclined to change their air conditioning settings due to social pressure or the impact of the sudden event, which can affect the air conditioning usage habits of surrounding residents. Therefore, the system introduces a social coupling term to simulate the conformity effect. Assume the intensity of social hot words in the group at a certain time period is... The social coupling correction factor It can be represented as: in, It is the Sigmoid activation function. To adjust the coefficient and control the intensity of social influence. When social topics such as "air conditioning" or "heat wave" surge on social media, The value will increase, leading more users to adjust their air conditioning settings, which in turn affects the fluctuation of the group's load.

[0060] As the number of users in a group increases, the impact of irrational behavior on the group's load may have a cumulative effect, necessitating the inclusion of a dynamic adjustment mechanism in the group-level load forecasting. Therefore, the system performs real-time weighted adjustments to the group load based on social signals and emotional fluctuations. Assume the adjustment factor for the group load within a certain time period is... The final output of the group load can be obtained through the following formula: in, It is a correction term based on the superposition of social emotional fluctuations, environmental factors, and irrational behavior. It can be estimated through regression analysis of historical behavioral data or other statistical models. In this way, the system can take into account the interaction effects between individuals in the group and dynamically adjust the predicted group load under different situations.

[0061] Another key task in group expansion and overlay is to consider behavioral differences between individuals. Even residents in the same area may have different air conditioning usage patterns due to differences in family structure, lifestyle, and price sensitivity. To better simulate this, the system uses a distributed model for training, where the behavior of each virtual agent is generated based on its unique feature data, thus reflecting the heterogeneity within the group. For example, younger users may tend to turn on the air conditioner more frequently when it is hot, while older users may choose to open windows for ventilation when the indoor temperature is high.

[0062] By employing this group expansion and overlay method, the system can generate representative group load forecast results based on the profiles and behavioral characteristics of different individuals. This approach not only considers the behavioral patterns of individual users but also effectively integrates irrational behaviors, social effects, and environmental influences at the group level, thereby providing more accurate and flexible support for electricity demand forecasting, load dispatch optimization, and demand response strategies.

[0063] In practical applications, the prediction of group load requires continuous calibration based on real-time data. The system can monitor the actual load data in the group in real time and compare it with the prediction results to perform error feedback correction. Whenever the system detects a deviation, it can update the model by adjusting the generation weights of the virtual agent or the irrationality correction factor to improve prediction accuracy and response flexibility. Ultimately, the group expansion and overlay stage not only provides group-level load prediction but also enables rapid adjustments in the event of emergencies, ensuring the efficiency and reliability of load scheduling.

[0064] 7. Uncertainty and Online Verification: The main objective of the uncertainty assessment and online calibration phase is to improve the accuracy and reliability of predictions by correcting model forecasts through real-time data feedback. In air conditioning load forecasting, due to changes in environmental conditions, the randomness of user behavior, and limitations of the model itself, the forecast results often contain a certain degree of uncertainty. To overcome this problem, the system employs an uncertainty assessment and online calibration strategy to dynamically adjust the forecast results to adapt to actual conditions.

[0065] First, during each load forecast, the system generates an air conditioning behavior script based on the input user profile and environmental scenario data, and calculates the corresponding power forecast value. However, there is often a difference between the actual load and the forecast load, stemming from factors such as changes in the external environment, irrational behavior, and equipment failure. To quantify this uncertainty, the system employs a diversity assessment method based on Monte Carlo sampling. In each forecast period, the system generates multiple different power curves to simulate the uncertainty of the forecast results. Specifically, it is assumed that the generated power curves are... (in Then, the prediction of the group load can be expressed as: in, The number of samples generated. This represents the average group load after diversity sampling. Using this method, the system can obtain a load forecast result that includes uncertainty.

[0066] To further improve forecast accuracy, the system incorporates an online calibration mechanism. The system triggers a calibration process whenever actual load data (from smart meters or area load monitoring) deviates significantly from the predicted load. Assuming a certain time period... The actual load is The predicted load is ,deviation It can be represented as: The system feeds this bias information back to the model and updates it through incremental learning or parameter tuning. Specifically, assuming the model's prediction function is... ,in Input features (including user profiles and environmental context). For the model parameters, the system updates them online by minimizing the following loss function: in, To calibrate the loss function, This represents the number of time steps during the calibration period. By optimizing this loss function, the system can fine-tune the model parameters according to changes in the actual load, making future predictions closer to reality.

[0067] In addition, the system incorporates Kalman filtering and other filtering methods to smooth and correct the prediction results. Kalman filtering is a recursive algorithm commonly used for state estimation of dynamic systems. It dynamically corrects the model output by taking a weighted average of the current predicted value and the actual observed value. Assume the system's state estimate is... The actual observed value is Then the update formula for the Kalman filter is: in, The Kalman gain is dynamically adjusted based on prediction errors and observation noise. In this way, the system can effectively eliminate short-term fluctuations and improve the stability of load forecasting.

[0068] During online calibration, the system not only adjusts model parameters based on real-time data but also considers the influence of historical data and long-term trends. To avoid overfitting short-term fluctuations, the system employs regularization to prevent the model from responding excessively to outliers during calibration. By adding a regularization term, the model can maintain a certain level of smoothness and generalization ability during updates. Assume the regularization term is... Then the final calibration loss function is: in, These are regularization coefficients used to balance the two terms in the loss function. The regularization term can be L2 regularization (e.g., ...). This prevents the model from becoming overly reliant on certain specific data points.

[0069] Through the above methods, the system can continuously optimize load forecasting results through real-time monitoring and adjustment. Whenever new actual data is collected, the system immediately triggers an online calibration process, dynamically updating model parameters based on actual deviations, thereby gradually improving future forecast accuracy. Ultimately, this not only enables precise adjustments to short-term load forecasts but also gradually improves the system's stability and reliability over long-term operation.

[0070] In summary, the core of the uncertainty and online calibration phase is to continuously optimize the model through a real-time feedback mechanism, enabling air conditioning load forecasting to better cope with uncertainties caused by environmental changes, fluctuations in user behavior, and other external factors. By introducing techniques such as diversity assessment, incremental learning, and Kalman filtering, the system can achieve efficient and dynamic load forecasting, improving the responsiveness of demand response and power dispatch.

[0071] 8. Output: At the end of the entire load forecasting process, the core task of the output phase is to generate accurate air conditioning load forecasts and provide them to downstream systems for scheduling and decision-making. The output is not just a single load value, but includes the forecast value and its uncertainty, risk assessment, and possible adjustment schemes, providing a scientific basis for subsequent power demand scheduling, load management, and demand response.

[0072] First, the system generates load forecast results for each time period, assuming the time period is... At each point in time The corresponding air conditioning load forecast is The prediction result is the final result after physical constraints and calibration of the power sequence output by the aforementioned model. The average value of the air conditioning load forecast can be expressed as: in, The number of samples generated. For the first The average load value for a given time period. This average value represents the expected group air conditioning load under given environmental and user behavior conditions.

[0073] In addition to the load forecast values, the system also needs to output the forecast uncertainty. To reflect the volatility of the air conditioning load forecast, the system generates multiple forecast samples and calculates the confidence interval of the load forecast using statistical methods. Assume the system generates multiple forecast values... (in The uncertainty can then be estimated by calculating the standard deviation and mean of these predicted values. in, For the first Average load forecast for the period The standard deviation for that period represents the range of confidence levels for the prediction. Based on the confidence level, the system can output the prediction interval, such as... This corresponds to a 95% confidence interval. This uncertainty information can help the scheduling system understand the volatility of the forecast and provide more comprehensive information for decision-making.

[0074] Furthermore, the system can output corresponding demand response suggestions or scheduling optimization schemes based on the predicted group load. For example, when a high load period is predicted, the system can adjust the air conditioning load allocation or suggest adjusting the air conditioning set temperature to achieve load smoothing and reduce grid pressure. In practical applications, the system can also calculate the demand response potential through a strategy model and generate predicted effects for different load management schemes. For example, assuming a high load warning occurs during a certain period, the system may suggest that residents reduce their air conditioning load by 10% by lowering the set temperature by 1°C. The calculation formula is as follows: in, For load changes, For temperature adjustment range, Power sensitivity related to temperature changes, This serves as the baseline power for the air conditioner. Based on this, the system can generate specific adjustment suggestions for the power dispatching system.

[0075] The system also needs to output risk assessments for each time period to quantify the reliability of load forecasting results. These risk assessments are based on the uncertainty of group load and feedback from real-time data, generating risk warnings under different scenarios. For example, when the load forecast exceeds a certain threshold, or when the forecast uncertainty is high, the system can mark it as a "high-risk" period and issue an early warning to remind the dispatching system to take preventative measures. The risk assessment can be calculated using the following formula: in, A risk index representing the load forecast results. The standard deviation of load forecasts This represents the average load forecast. A higher risk value indicates lower reliability of the load forecast for that period, potentially requiring stronger dispatch intervention measures.

[0076] Finally, the system will also output time-of-use pricing and demand response potential. Assuming a certain time period... The predicted air conditioning load is The electricity price during that period was The system will calculate the cost-benefit of that period: Based on electricity prices and load forecasts, the system can provide users with real-time energy-saving suggestions. For example, during periods of high electricity prices, the system may suggest that residents appropriately increase their air conditioner settings or delay the start-up time to reduce load and electricity costs. This strategy can combine user behavior preferences with social electricity price signals to provide intelligent support for demand response.

[0077] In summary, the output phase goes beyond simply providing a single load forecast. It offers comprehensive data support for power dispatch and demand response by providing multi-dimensional information such as the load average, standard deviation, risk assessment, demand response recommendations, and electricity price analysis. By comprehensively considering forecast uncertainty and risk assessment, the system can better address uncertainties in actual operation and provide accurate and flexible decision support for power management departments.

[0078] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. < / irr-high>

Claims

1. A large-scale model prediction method for air conditioning load considering external environment and irrational behavior, characterized in that, Includes the following steps: S1: Collect residents' basic information and generate static profile vectors. If there is historical natural language text data, extract semantic vectors. After fusion, form the final individual profile vector. After logical verification, large language model completion and clustering, output numerical profile vectors and semantic profile summary text. S2, collect environmental data within the prediction time window, construct environmental vector sequences and scenario prompt text, and extract behavioral fragments similar to the current scenario; S3, the generated numerical profile vector, semantic profile summary text, temporal environment vector and contextual prompt text are input into the large language model to generate an air conditioning control script including timestamp, operation type and parameters; S4 parses the air conditioning control script into structured operation tuples, combines the external temperature, set temperature, wind speed and mode, calculates the instantaneous power of the air conditioner, and integrates the irrational coefficient and social influence to form an initial power sequence; S5, based on thermal inertia and equipment limitations, performs physical consistency correction on the initial power curve, constrains the start-stop frequency and maximum power fluctuation, and forms a corrected power curve; S6: Sample resident profiles to generate virtual resident agents, overlay power output and introduce social influence to generate predicted values ​​of group air conditioning load; S7 uses Monte Carlo sampling to construct an uncertainty prediction set and performs online learning and short-time error correction when prediction bias occurs; S8 outputs the mean, standard deviation, and confidence interval of the predicted air conditioning load for the group, generating risk assessment and adjustment recommendations.

2. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 1, characterized in that, S1 includes: S11, collect basic information for each resident, including age. ,income ,Profession Housing type Family structure and historical energy consumption statistics Generate resident attribute vectors , is represented as: ； S12, the resident attribute vector is input into the feature embedding network for encoding, generating a static profile vector. , is represented as: ； in, Encoding function for static images; S13, If residents have historical natural language text data, use a semantic encoder to extract their semantic vectors. , is represented as: ； in, For user-related text collections, For semantic text encoding functions; S14 concatenates the static portrait vector with the semantic vector and maps them to a unified semantic feature space to form the final individual portrait vector. , is represented as: ； S15, the final individual profile vector undergoes logical relationship verification and statistical anomaly detection. Outlier detection is performed on the feature dimensions using standard deviation normalization, as shown below: ； when When this occurs, a correction or marking as an abnormal user is triggered; S16, for users with missing or abnormal text, a large language model is invoked to generate complete text, and clustering matching is performed based on the embedding results and historical image vectors in the image database. Cosine similarity is used to define the semantic similarity of the images, expressed as: ； when If the image deviates significantly from the known group, it indicates that manual verification or automatic re-inference is required. The similarity threshold; S17 outputs numerical portrait vectors and semantic summary text.

3. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 2, characterized in that, S2 includes: S21, Set the prediction time window Extracting external temperature data from multi-source databases relative humidity Wind speed Solar radiation intensity Air Quality Index Time-of-use electricity pricing Demand response signals Event tag sequence and frequency of social keywords Environmental feature vectors are constructed from the environmental data, including those from the environment. , is represented as: ； S22, for environmental feature vectors Perform normalization and smoothing operations for each feature dimension. Using sliding window mean filtering and normalization transformation, it can be expressed as: ； ； in, To smooth the window width, , These are the historical mean and standard deviation; S23, All normalized environmental feature vectors are combined into an environmental vector sequence, represented as: ； S24, Use the semantic mapping module to convert the environment vector sequence into semantic contextual cue text. ; S25 introduces a retrieval enhancement generation mechanism based on a large language model, which calculates query vectors through a semantic indexing model. It then retrieves historical behavior segments similar to the current environmental context from the constructed environmental corpus, calculates similarity using vector cosine similarity, and selects the segments with the highest similarity. The results of a similar scenario search are represented as follows: ； S26, Output the numerical environment vector sequence and semantic contextual cue text .

4. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 3, characterized in that, S3 includes: S31, for numerical image vectors Semantic summary text Numericalized environment vector sequence and semantic contextual cue text This concatenation is used to drive a large language model to generate event sequences. Each event consists of a triple, represented as: ； in, For timestamps, As an action type, For parameters; S32, the script event sequence is generated through maximum likelihood estimation, represented as: ； in, These are learnable parameters; S33, using instruction fine-tuning to minimize standard cross-entropy for supervised training, is expressed as: ； S34 introduces an irrational behavior mapping function to define the irrationality coefficient of an individual's response to the environment, expressed as: ； in, This is a statistical summary of the environment vectors within the time window. As an indicator of sentiment or conformity in the intensity of social trending words, For the Sigmoid function; The temperature parameter is set as follows: ； S35 introduces feasibility and vocabulary constraints, defining a feasibility penalty function as follows: ； in, A large penalty coefficient, For indicator functions; S36, apply vocabulary sparsity regularization to the generated parameters and introduce a temperature coverage regularization term to control and adjust the sparsity, expressed as: ； ； in, For a set of sliding time windows, The minimum number of events threshold; S37, through joint optimization of behavior generation loss, temperature coverage regularization, and physical consistency loss, is expressed as: ； ； S38 defines a timestamp position error function to control the accuracy of event generation and outputs the script sequence and its corresponding execution confidence, expressed as: 。 5. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 4, characterized in that, S4 includes: S41, perform semantic parsing and structured processing on the natural language segments in the air conditioning control script, generating a sequence including timestamps. Operation type Parameter values and operating mode The Quadruple Event , is represented as: ； ； in, This is a mapping function from language embeddings to parameter differences; S42, based on the operation type, set parameters, and environmental conditions, calculates the instantaneous power of the air conditioner, expressed as: ； in, This refers to standby power. This is the power-on indicator function. These are the equipment's characteristic parameters; S43, using the pattern mapping function The modified energy efficiency parameter set is expressed as follows: ； S44 introduces a thermal balance differential equation to model the dynamic response of air conditioning operation to indoor temperature, expressed as: ； ； S45, introducing an irrational correction coefficient. And the power is corrected once, expressed as: ； S46, introduces a sliding fusion mechanism to compare power with the average power of the local population. The power is fused to form a smoothed power estimate, expressed as: ； in, The social coupling coefficient; S47, perform a sliding integral on the power sequence to obtain the final smoothed power output, and apply boundary constraints, as follows: ； ； S48 outputs the final time series power data, represented as follows: 。 6. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 5, characterized in that, S5 includes: S51 introduces a simplified first-order heat balance equation to simulate the change process of indoor temperature, expressed as: ； in, For the heat capacity of the building, For thermal resistance, This refers to the cooling efficiency factor of the air conditioner. This refers to the instantaneous power of the air conditioner. S52 introduces a physical consistency correction term to correct the power response where the current power and indoor temperature deviation directions are inconsistent, expressed as: ； in, This is the power-on indicator. To prevent small constants from being divided by zero; S53 sets the maximum and minimum operating power for the air conditioning equipment, expressed as: ； S54, Set minimum start-stop interval With maximum adjustment range , is represented as: ； S55, the power series is smoothed using a moving average method, expressed as: ； in, To adjust the sliding window size; S56 introduces physical consistency and device constraints to correct the power sequence, correct unreasonable temperature control response behavior, and generate a corrected power curve.

7. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 6, characterized in that, S6 includes: S61 samples profile information from different users, generates virtual resident agents through a large language model, and outputs the air conditioning load of the virtual agents. ; S62, perform a weighted summation of the air conditioning loads of all virtual agents to form a group load estimate. , is represented as: ； in, as an agent The weight, The number of agents in the group; S63 introduces a social keyword frequency correction factor to construct social coupling. , is represented as: ； in, It is the Sigmoid activation function. This is the adjustment coefficient; S64 introduces a correction factor for group load. The final predicted population load is constructed by incorporating a social coupling correction factor, and is expressed as: 。 8. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 7, characterized in that, S7 includes: S71 will extract user profiles and environmental data, construct a perturbation distribution script, perform Monte Carlo sampling from it, and generate multiple sets of sample power curves. And calculate the population average predicted power. , is represented as: ； in, The number of samples generated; S72, at each actual sampling time Calculate the predicted value Compared with the actual measured value The difference , is represented as: ； Based on difference Define the calibration loss function as follows: ； S73 introduces Kalman filtering to dynamically adjust the predicted power, expressed as: ； S74, introduces a regularization term. To avoid overfitting, the final calibration loss function is expressed as: ； in, This is the regularization coefficient.

9. The large-scale air conditioning load prediction method considering external environment and irrational behavior according to claim 8, characterized in that, S8 includes: S81, in each time period The above generates predicted air conditioning load values ​​from multiple samples. The average value is calculated to reflect the overall load trend at that moment, and is expressed as: ； in, The number of samples generated; S82, calculate the standard deviation of the air conditioning load forecast results for each time period, and construct the confidence interval, as follows: ； ； ； in, For the first Forecasted average air conditioning load for the time period; S83, taking into account the sensitivity to temperature and load changes, generates load regulation recommendations, expressed as follows: ； in, For load changes, For temperature adjustment range, Power sensitivity related to temperature changes, This is the base power of the air conditioner; S84 calculates load risk indicators using the standard deviation and average value of air conditioning load forecasts. , is represented as: ； S85, based on the predicted air conditioning load and time-of-use electricity pricing, calculates the cost-effectiveness for each time period, expressed as: ； S86 ultimately outputs the predicted air conditioning load, standard deviation, confidence interval, load risk index, cost-effectiveness, and adjustment recommendations.

Images