A Chiller Load Optimization Algorithm for Central Air Conditioning Chiller Units Based on a Large Language Model

Abstract

This invention relates to the field of intelligent optimization technology, specifically to a chiller load optimization algorithm for central air conditioning chiller units based on a large language model. The algorithm includes the following steps: extracting chiller load terms from scheduling requests through word vector parsing; calculating the offset sequence and using a sliding window to identify load switching points and divide the control stages; generating control instructions by filtering equipment actions based on TF-IDF; extracting target parameters using dependency parsing to filter conflicting terms and form a parameter chain; and calculating pump frequency using linear mapping to generate execution instructions. In this invention, the chiller load offset sequence is dynamically generated through semantic parsing of natural language scheduling requests and word vector distance; the control stages are precisely divided by identifying switching points using a sliding window; and parameter chains are constructed using co-occurrence frequency and dependency analysis to filter conflicting instructions. This improves response accuracy and strategy matching, strengthens the capture of implicit intent, reduces data dependence and experience bias, ensures the safety of multi-unit collaborative optimization, and reduces the risk of energy efficiency fluctuations.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology, and in particular to a chiller load optimization algorithm for central air conditioning chiller units based on a large language model. Background Technology

[0002] The field of intelligent optimization technology refers to a collection of related technologies that utilize mathematical modeling, operations research methods, and artificial intelligence algorithms to assess the state, adjust parameters, and control processes of complex systems, aiming to achieve comprehensive goals such as rational resource allocation, improved operational efficiency, and optimized energy consumption. Core aspects of this field include problem modeling, optimization goal setting, constraint handling, algorithm design and solving, and result feedback mechanisms. Its applications are widely distributed across various practical scenarios, including energy system scheduling, traffic route planning, manufacturing process optimization, logistics and warehousing management, and energy-saving control of air conditioning systems. It emphasizes the formulation and execution of strategies through optimization methods within complex systems.

[0003] Traditional central air conditioning chiller load optimization algorithms address the issue of cooling load allocation in building centralized cooling systems when multiple chillers operate simultaneously. These algorithms employ fixed load allocation rules, empirical coefficient allocation methods, or heuristic search strategies to distribute the load. Based on indicators such as the unit's start-up / shutdown status, operating frequency, cooling demand, and historical load data, these methods determine the load allocation value for each unit through simple calculation formulas or pre-set empirical models to achieve a basic balance in the system's cooling supply.

[0004] Traditional methods rely on fixed load allocation rules and empirical coefficients. Their static allocation logic is difficult to match dynamic changes in cooling demand. Fixed rules lack the ability to parse unstructured information in text scheduling requests, resulting in incomplete extraction of load intensity offset features. Empirical coefficient allocation methods are limited by the assumption of linear correlation of historical load data and cannot capture the semantic correlation of real-time control intentions. Heuristic search strategies ignore the nonlinear change characteristics of continuous offsets when judging load switching, causing delays or missegments in stage division. The equipment-action combination screening does not integrate the dynamic verification mechanism of co-occurrence frequency and physical constraints, which easily generates instructions that exceed the safety boundary. Ultimately, this leads to insufficient real-time performance and robustness of the chiller load allocation strategy. Summary of the Invention

[0005] To address the shortcomings of traditional methods that rely on fixed load allocation rules and empirical coefficients, which suffer from static allocation logic that struggles to match dynamic cooling demand changes, fixed rules lack the ability to parse unstructured information in text scheduling requests leading to incomplete extraction of load intensity offset features, empirical coefficient allocation methods are limited by the linear correlation assumption of historical load data and cannot capture the semantic correlation of real-time control intentions, heuristic search strategies ignore the nonlinear variation characteristics of continuous offsets when judging load switching, causing delays or missegments in stage division, and the lack of a dynamic verification mechanism that integrates co-occurrence frequency and physical constraints in equipment-action combination screening, which easily generates instructions exceeding safety boundaries, ultimately resulting in insufficient real-time performance and robustness of chiller load allocation strategies, this invention provides a chiller load optimization algorithm for central air conditioning chiller units based on a large language model. The technical solution is as follows:

[0006] On the one hand, a chiller load optimization algorithm for central air conditioning chiller units based on a large language model is provided, the algorithm including:

[0007] S1: Semantically analyze the natural language scheduling request text using a word vector model, extract the set of terms describing the operating status of the chiller unit, and perform cooling load intensity offset calculation on adjacent terms based on word vector spatial distance calculation to generate a cooling load intensity offset sequence.

[0008] S2: Call the cooling load intensity offset sequence and iterate through it using a sliding window mechanism. When the absolute value change rate of three consecutive offsets exceeds the chiller start-up and shutdown threshold, mark it as a load switching point. Segment the text according to the switching point position to obtain the cooling load control stage.

[0009] S3: Invoke the cooling load control stage, based on the TF-IDF algorithm, count the co-occurrence frequency of chiller unit equipment names and control actions, filter equipment-action combinations with frequency values ​​greater than the chiller unit action response threshold, and generate a chiller unit control instruction set;

[0010] S4: Invoke the chiller unit control instruction set, extract the target parameter items directly related to the chiller load in the scheduling request through dependency parsing, sort the parameters that have not been negated, combine the physical constraints of the equipment-action combination, filter the instruction items that conflict with the safe operating range of the chiller unit, and form the chiller load target parameter chain.

[0011] As a further aspect of the present invention, the cooling load intensity offset sequence includes word vector offset, load intensity change rate and timestamp sequence; the cooling load control stage includes a set of switching time points, time period division identifier and associated temperature parameter group; the chiller unit control instruction set includes instruction coding standard, operating parameter range and validity verification identifier; and the chiller load target parameter chain includes parameter priority sequence, constraint relationship network and safe operation threshold dataset.

[0012] As a further aspect of the present invention, the specific steps of S1 include:

[0013] S101: Semantic parsing of natural language scheduling request text is performed through word vector model. After extracting keywords from the text, the associated terms of chiller unit operation status are identified. The model embedding parameters are called to convert each term into a vector representation and generate a sequence of operation status term vectors.

[0014] S102: Based on the word vector sequence of the running state, extract the vector combination of adjacent words, use Euclidean distance as the measure of positional offset between vectors, calculate the spatial distance between word vectors pairwise in sequence, organize the results into a linear distance sequence, and obtain the semantic distance sequence between words;

[0015] S103: Call the semantic distance sequence between the terms, match the cooling load variation level corresponding to the differential distance value according to the vector change benchmark of the chiller unit's operating condition, map the position difference to the specific intensity range according to the distance sequence order, fit the numerical correspondence between the index sequence and the intensity quantity, and generate the cooling load intensity offset sequence.

[0016] The operating condition state vector benchmark is a set of typical load vectors formed by clustering in the original cold load operating data.

[0017] As a further aspect of the present invention, the specific steps of S2 include:

[0018] S201: Call the cooling load intensity offset sequence, set a sliding window mechanism to traverse it, extract the absolute value of the offset within the window, calculate the rate of change of two adjacent sets of offset values ​​and organize them into a sequence to generate a continuous offset rate of change sequence.

[0019] S202: Based on the continuous offset change rate sequence, combined with the chiller unit start-up and shutdown threshold, determine whether multiple consecutive values ​​exceed the limit simultaneously, record the corresponding index position and map it to the original time point to obtain the cooling load switching time point sequence.

[0020] The chiller start-stop threshold is a critical standard for judging whether the load change has reached the critical point of needing to start or stop the unit. It is used to identify the load change trend and guide the control decision. It is set based on the empirical value of the chiller's response to load change during the test phase, and is 10%~15% of the cooling load change rate.

[0021] S203: Based on the cooling load switching time point sequence, call the cooling load intensity offset sequence to divide it in time order, use the position in the time point sequence as the dividing boundary to extract the interval offset data and mark the segment number to establish the cooling load control stage.

[0022] As a further aspect of the present invention, the specific steps of S3 include:

[0023] S301: Based on the control log text recorded during the cooling load control phase, the TF-IDF algorithm is used to calculate the weight value of the terms in the control statement, extract the terms including the name of the chiller unit equipment and the control action, count the number of times the equipment and action combination co-occurs in the text, and obtain the equipment action co-occurrence frequency value.

[0024] The co-occurrence frequency value of the equipment action represents the frequency at which a certain chiller unit equipment and a certain control action appear together in the control log, and is used to measure the probability of the combination occurring.

[0025] S302: Based on the co-occurrence frequency value of the equipment actions, and using the chiller unit action response threshold as a benchmark, determine whether the frequency value of each group of equipment and actions meets the response standard, filter out combinations that do not meet the benchmark, sort out the combinations that meet the conditions according to equipment classification, and obtain the response combination frequency set.

[0026] The chiller unit action response threshold is the minimum frequency standard for judging whether the equipment-action combination has responsive significance; combinations below this value will be screened out.

[0027] S303: Call the device-action combination in the response combination frequency set, organize the combination items according to the chiller unit number order, set the standard format according to the action description and perform structural coding, integrate the action instructions under the device, and generate the chiller unit control instruction set.

[0028] As a further aspect of the present invention, the specific steps of S4 include:

[0029] S401: Invoke the chiller unit control instruction set, based on the syntactic dependency relationship structure tree in the scheduling request, parse the subject-verb-object relationship and modifiers, identify the parameter phrases associated with the chiller load, remove negative terms, sort the parameters according to syntactic weights, and generate the load target parameter sequence value;

[0030] S402: Based on the load target parameter sequence value, for the device-action combination configuration, compare the adaptability of the sequence parameters with the controller action items, detect the compatibility of the response and execution instructions of the feedback channel, eliminate conflicting instructions, and obtain the parameter and action coordination matching rate value.

[0031] S403: Call the parameter and action coordination matching rate value, and perform interval logic judgment on the distribution value of the parameter item in the target action range according to the upper and lower limit thresholds of the chiller unit operation. Eliminate parameters in non-control sections, retain the instruction parameter combination that meets the safety constraints, and generate the chiller load target parameter chain.

[0032] As a further aspect of the present invention, the load target parameter sequence value is an ordered set of parameters formed by extracting control parameters associated with the chiller load based on syntactic dependency analysis and sorting them according to semantic weights.

[0033] The parameter-action coordination matching rate value is an adaptation assessment result that measures the degree of matching between parameters and equipment control actions in terms of logic, execution, and feedback.

[0034] The upper and lower operating thresholds of the chiller unit are multi-dimensional operating boundary parameters of load, temperature, and pressure that limit the safe operating range of the chiller unit.

[0035] As a further aspect of the present invention, it includes:

[0036] S5: Based on the parameter item ranked first in the chiller load target parameter chain, call the corresponding equipment-action combination in the chiller unit control instruction set, linearly map the parameter item value with the equipment action range, calculate the chilled pump frequency adjustment percentage value, and generate optimized execution instructions.

[0037] The optimized execution instructions include range conversion coefficients, frequency adjustment ranges, and timing control time nodes.

[0038] As a further aspect of the present invention, the specific steps of S5 include:

[0039] S501: Based on the load parameter item ranked first in the chiller load target parameter chain, retrieve the equipment-action combination item in the chiller unit control command set, screen the chilled pump as the frequency regulation target device, extract its frequency range boundary and parameter item value range, establish a linear correspondence function and calculate the mapping value, and obtain the frequency response ratio value.

[0040] S502: Call the frequency response ratio value, read the current initial frequency value according to the device frequency adjustment threshold range, construct an adjustment function by combining the frequency starting point and the ratio value, calculate the direction and amplitude of frequency change, and generate a pump speed gain value;

[0041] S503: Based on the pump speed gain value and the field structure constraint rules in the refrigeration pump control command template, integrate the control cycle timestamp and device identifier, encapsulate the frequency change value to construct a standard data frame, write it into the device execution interface data channel, and generate an optimized execution instruction.

[0042] As a further aspect of the present invention, the frequency response ratio represents the linear mapping ratio of the current load parameter change within the frequency range of the chilled pump;

[0043] The pump speed gain value represents the actual frequency increase or decrease of the refrigeration pump calculated based on the response ratio value;

[0044] The adjustment function is a mathematical model constructed based on the response ratio and the current frequency to calculate the target frequency.

[0045] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0046] Based on semantic parsing and word vector spatial distance calculation of natural language scheduling requests, a cold load intensity offset sequence is dynamically generated. The sliding window mechanism identifies load switching points where the rate of change of continuous offset exceeds the threshold, achieving precise segmentation of the cold load control stage. Combined with equipment-action co-occurrence frequency screening and dependency parsing, a target parameter chain is constructed and conflicting instructions are filtered, improving the dynamic response accuracy and strategy matching of cold load allocation. This enhances the ability to capture implicit control intentions in text descriptions, reduces reliance on historical data and human experience bias, ensures that the multi-unit collaborative optimization process complies with physical safety constraints, and reduces the risk of system energy efficiency fluctuations. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0048] Figure 2 This is a detailed flowchart of S1 of the present invention;

[0049] Figure 3 This is a detailed flowchart of the S2 process of the present invention;

[0050] Figure 4 This is a detailed flowchart of the S3 process of the present invention;

[0051] Figure 5 This is a detailed flowchart of the S4 process of the present invention;

[0052] Figure 6 This is a detailed flowchart of S5 of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0054] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0055] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0056] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0057] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0058] Please see Figure 1 This invention provides a chiller load optimization algorithm for central air conditioning chiller units based on a large language model. The processing flow of this algorithm may include the following steps:

[0059] S1: Semantically analyze the natural language scheduling request text using a word vector model, extract the set of terms describing the operating status of the chiller unit, and perform cooling load intensity offset calculation on adjacent terms based on word vector spatial distance calculation to generate a cooling load intensity offset sequence.

[0060] S2: Call the cooling load intensity offset sequence and iterate through it using a sliding window mechanism. When the absolute value change rate of three consecutive offsets exceeds the chiller start-up and shutdown threshold, mark it as a load switching point. Segment the text according to the switching point position to obtain the cooling load control stage.

[0061] S3: Invoke the cooling load control stage, based on the TF-IDF algorithm, count the co-occurrence frequency of chiller unit equipment names and control actions, filter equipment-action combinations with frequency values ​​greater than the chiller unit action response threshold, and generate chiller unit control instruction set;

[0062] S4: Call the chiller unit control instruction set, extract the target parameter items directly related to the chiller load in the scheduling request through dependency parsing, sort the parameters that have not been negated, combine the physical constraints of the equipment-action combination, filter the instruction items that conflict with the safe operating range of the chiller unit, and form the chiller load target parameter chain.

[0063] S5: Based on the parameter item ranked first in the chiller load target parameter chain, call the corresponding equipment-action combination in the chiller unit control instruction set, linearly map the parameter item value with the equipment action range, calculate the chilled pump frequency adjustment percentage value, and generate optimized execution instructions.

[0064] The cooling load intensity offset sequence includes word vector offset, load intensity change rate and timestamp sequence; the cooling load control stage includes switching time point set, time period division identifier and associated temperature parameter group; the chiller unit control instruction set includes instruction coding standard, operating parameter range and validity verification identifier; the chiller load target parameter chain includes parameter priority sequence, constraint relationship network and safe operation threshold dataset; the optimization execution instruction includes range conversion coefficient, frequency adjustment range and timing control time node.

[0065] Specifically, such as Figure 2 As shown, the specific steps of S1 are as follows:

[0066] S101: Semantic parsing of natural language scheduling request text is performed through word vector model. After extracting keywords from the text, the associated terms of chiller unit operation status are identified. The model embedding parameters are called to convert each term into a vector representation and generate a sequence of operation status term vectors.

[0067] The process of semantically parsing natural language scheduling request text using a word vector model first requires receiving the user-submitted scheduling request statement, such as "reduce the operating load of chiller unit 5" or "switch chiller unit 3 to energy-saving mode." This text is then input into a pre-trained Chinese word vector model, such as the Word2Vec model. This model is trained using corpora from central air conditioning system operation logs, equipment alarm records, and historical work order data, with a training scale of no less than 500,000 sentences. Each word's vector dimension is set to 128 dimensions, and a Skip-gram architecture is used during training, with a window size of 5. After receiving the input text, the model performs word segmentation. For example, the sentence "reduce the operating load of chiller unit 5" is broken down into "reduce," "number 5," "chiller unit," "operate," and "load." The model then looks up the corresponding word vector for each term; for example, the vector for "reduce" is... The remaining terms can also obtain a similar 128-dimensional vector representation, ultimately forming a term vector sequence of length 5, i.e. The process then identifies terms associated with the chiller unit's operating status, such as "operation," "load," "energy saving," and "switching," and constructs a state-related word vector subsequence. Taking "switching to energy-saving mode" as an example, the key terms are "switching," "energy saving," and "mode." After extracting their word vector sequences, another state representation subsequence can be formed. During the model embedding parameter call process, it is necessary to ensure that the terms used have training records in the word vector model. If some terms are new words or technical terms, such as "COP coefficient" or "secondary pump," they need to be completed through word concatenation or average embedding. For example, for the phrase "adjusting the outlet water temperature," if the term "outlet water" does not have a corresponding embedding vector in the model dictionary, the average of the word vectors for "outlet" and "water" can be taken, or the concatenation can be performed to reduce dimensionality and generate a synthetic term vector. This process ultimately outputs the operating status term vector sequence for each request text, laying the input foundation for subsequent vector combination and semantic distance calculation. To verify the processing effect, the following actual request sentences can be used for processing, such as "Increase the outlet water temperature of Unit 3 to 7℃", "Stop the operation of Unit 1 for 30 minutes", and "Unit 5 has too low energy efficiency". The term extraction and vector transformation are shown in the table below:

[0068] Table 1: Sequence of term vectors in the scheduling request text (partial display)

[0069]

[0070] As shown in Table 1, the scheduling text can be successfully converted into a sequence of 128-dimensional vectors. For terms like "too low," which have a qualitative descriptive nature, the system quantifies them using actual operational data. For example, when the chiller unit's Coefficient of Performance (COP) is below 3.2, it is considered "too low." The "too low" vector in the corresponding vector sequence will be close to the COP-related clustering label vector, thus supporting semantic alignment. The final generated term vector sequence provides the basis for calculating the semantic offset in subsequent steps.

[0071] S102: Based on the sequence of word vectors in the running state, extract the vector combination of adjacent words, use Euclidean distance as the measure of positional offset between vectors, calculate the spatial distance between each pair of word vectors in sequence, organize the results into a linear distance sequence, and obtain the semantic distance sequence between words;

[0072] Based on the sequence of term vectors in the running state, we first need to extract adjacent vector combinations pairwise according to the order of the terms in the original text. For example, for the vector sequence... Combine them in pairs Each vector pair is used for subsequent spatial distance calculations. The Euclidean distance method is used to measure the distance between the two vectors. The specific calculation formula is as follows:

[0073] ;

[0074] in Indicates the first The term in the first The vector components are 4-dimensional. For example, the first 4 dimensions of the vectors "improvement" and "number three" are as follows:

[0075] and ;

[0076] Then its first 4-dimensional Euclidean distance is:

[0077] ;

[0078] Perform the same operation on all 128 dimensions to obtain the complete distance values. Arrange the spatial distances of adjacent terms in order to form a linear distance sequence. This sequence reflects the relative changes between terms in the semantic space, where the value range is determined based on the statistical results of the training data. Between terms, if the distance between a pair of terms is less than 0.5, they are considered to have a high semantic correlation; if the distance is greater than 1.5, they are considered semantically broken regions. In practical applications, if the distance sequence generated by processing a statement such as "reduce the operating load of chiller unit No. 5" is... This indicates a strong semantic coherence between "operation" and "load," while a certain semantic shift exists between "chiller unit" and "operation." At this stage, the system needs to define a precision range for each distance value, store it to three decimal places, and establish a standardized mapping to normalize the distance values ​​to the specified interval. To match the mapping requirements of subsequent cooling load variation levels, the normalization method adopts a maximum-minimum scaling approach, specifically:

[0079] ;

[0080] in , Therefore, the normalized result corresponding to the original value of 0.45 is 0.225. After this step, the semantic distance sequence is obtained. This serves as an important reference for further determining the level of cooling load variation.

[0081] S103: Call the semantic distance sequence between terms, match the cooling load variation level corresponding to the differential distance value according to the vector change benchmark of the chiller unit's operating condition, map the position difference to the specific intensity range according to the distance sequence order, fit the numerical correspondence between the index sequence and the intensity quantity, and generate the cooling load intensity offset sequence.

[0082] The operating condition state vector benchmark is a set of typical load vectors formed by clustering in the original cold load operating data;

[0083] After calling the semantic distance sequence, each distance value needs to be classified and compared according to the pre-constructed chiller unit operating condition state vector benchmark. This benchmark set is derived from the cluster analysis of historical operating data of the chiller unit. For example, K-Means clustering is performed on the average daily operating load data within one year, with the number of clusters set to 5, representing five typical operating states: extremely low load, low load, medium load, high load, and extremely high load. Each state is defined with a 128-dimensional vector center point. That is, the reference set of the working condition state vector is During the comparison process, each semantic distance value The distance value will be used to calculate its corresponding load intensity offset level. The system maps the distance value to the level space according to a pre-set threshold interval, for example, the interval is set as follows: This corresponds to "no significant change". This corresponds to "slight changes". This corresponds to "moderate change". This corresponds to "drastic change". This corresponds to "mutation". (Based on semantic distance sequences) For example, the corresponding change levels are "slight change", "slight change", "drastic change" and "no significant change", and the system constructs a set of change level codes accordingly. This represents five categories of variation intensity, with each distance value corresponding to a variation intensity code, ultimately yielding a cooling load intensity offset sequence. To further develop a structured index relationship, a mapping table is established between the index position numbers of the original semantic distance sequence (e.g., 1, 2, 3, 4) and the variable intensity code. A numerical function is then established using polynomial fitting or linear interpolation. ,in For index position, The corresponding intensity values ​​are used to generate a cooling load intensity offset sequence, which serves as the basic quantitative indicator for judging load changes when the chiller unit is dispatched.

[0084] Specifically, such as Figure 3 As shown, the specific steps of S2 are as follows:

[0085] S201: Call the cooling load intensity offset sequence, set a sliding window mechanism to traverse it, extract the absolute value of the offset within the window, calculate the rate of change of two adjacent sets of offset values ​​and organize them into a sequence, and generate a continuous offset rate of change sequence.

[0086] After calling the cooling load intensity offset sequence, first set the sliding window length, for example, setting the window length to 3. This means that each iteration selects three adjacent offset values ​​to form a set of data. If the cooling load intensity offset sequence is... The result of window traversal is , , , In each set of data, extract two adjacent offset values ​​and calculate the absolute value of their difference. Then, using the previous value as the denominator, convert the absolute value into a rate of change. For example, in the first set, calculate... , If the denominator is 0, skip that term or set it to 0 for processing. The rate of change forms a continuous rate of change sequence. Execute the second group. When processing, the rate of change for the first group was 2, and the rate of change for the second group was 3, respectively by... , By extension, by performing rate-of-change processing on each group of sliding window results and storing the rate-of-change results in an array, the final continuous offset rate-of-change sequence is obtained. Each pair of change rates is calculated from adjacent offset values. During the traversal, the change rate of each group needs to be indirectly correlated with the first and second values ​​of the next group to construct a time-series trend. In this process, large differences in offset values ​​will lead to a significant increase in the change rate; if the offset values ​​remain unchanged for a long time, the continuous change rate will approach 0. To avoid error interference, the original offset sequence needs to be preprocessed before traversal, such as removing abrupt changes or invalid points, to ensure that the data within the window is continuous and valid. The offset sequence is set through an instance. When the sliding window length is 3, a total of 5 windows are generated, namely: , , , , The rates of change were calculated as follows: , , , , Generate the final continuous offset change rate sequence As shown in Table 2.

[0087] Table 2: Series of Change Rates in Cooling Load Intensity Deviation

[0088]

[0089] Table 2 shows the results of the rate of change calculation for the cold load offset sequence using a sliding window length of 3.

[0090] S202: Based on the continuous offset change rate sequence, combined with the chiller unit start-up and shutdown threshold, determine whether multiple consecutive values ​​exceed the limit simultaneously, record the corresponding index position and map it to the original time point to obtain the cooling load switching time point sequence;

[0091] The chiller start-up and shutdown threshold is a critical standard for judging whether the load change has reached the point where the unit needs to be started or stopped. It is used to identify the load change trend and guide the control decision. It is set based on the empirical value of the chiller's response to load change during the testing phase, and is 10%~15% of the cooling load change rate.

[0092] Obtain the continuous offset change rate sequence:

[0093] ;

[0094] The system needs to identify whether load changes have exceeded acceptable limits based on the chiller unit start-up and shutdown thresholds. The start-up and shutdown thresholds are set at 10%~15%, meaning that when the change rate is greater than 0.10~0.15, it is determined that the load change trend at that moment requires unit control. During the judgment process, the system compares each change rate with the upper and lower limits of the set thresholds. For example, if the lower limit is 0.10 and the upper limit is 0.15, then values ​​of 0.5, 1.0, 2.0, etc., all exceed the upper limit and are considered over-limit points. The corresponding index position information is recorded. For example, in the above change rate sequence, index positions 02 and 69 are both over-limit segments. The time point corresponding to the original sequence is the start time of the offset sequence plus the number of window steps. That is, if the original sequence has a time step of 1 minute, then the time point corresponding to index position 02 is the 13th minute, and 69 corresponds to the 710th minute. This moment is used as the switching boundary point of the cooling load change. To improve accuracy, the minimum length of consecutive exceedances is further determined. For example, a valid switching segment is considered if the rate of change is greater than the upper limit for at least two consecutive time points. If the rate of change at a certain point is large but there is no continuity before or after it, it is discarded. During execution, a cumulative discrimination method is used to count consecutive high values. After reaching the standard, the first and last indices are recorded and mapped to time points. Taking this example, consecutive segments such as [1.0, 1.0, 1.0] and [2.0, 2.0, 1.0] meet the conditions, and the final output switching time points are the 1st minute and the 7th minute. This judgment is based on a threshold of 10%~15%. To make the setting more reasonable and specific, historical start-up and shutdown data of the chiller unit are selected. The average load change rate is calculated to be 12.8% through difference calculation. The upper limit of the start-up and shutdown threshold is set to 15%, and the lower limit is set to 10%. Refer to the following actual example: If the chiller unit load changes from 450RT to 510RT within 5 minutes, the change rate is:

[0095] ;

[0096] If the value is between 10% and 15%, it indicates a critical switching situation. If it occurs twice in a row, the switching is considered effective, the original time point is mapped, and it is saved as a switching time point sequence.

[0097] S203: Based on the cooling load switching time point sequence, the cooling load intensity offset sequence is called and divided in chronological order. The position in the time point sequence is used as the dividing boundary to extract the interval offset data and mark the segment number to establish the cooling load control stage.

[0098] After obtaining the cooling load switching time sequence, for example, if the switching time is at minute 1, minute 7, and minute 12, call the complete cooling load intensity offset sequence:

[0099] ;

[0100] Using the aforementioned time points as segment boundaries, the corresponding time period offset data is extracted to construct multiple control phase segments. Specifically, the first phase begins at minute 1 and ends at minute 7; the second phase begins at minute 7 and ends at minute 12; and the third phase follows. Within each segment, the time series and offset values ​​are retained, correspondingly labeled as phases P1, P2, and P3. An independent index sequence is established for the offset values ​​within each phase. For example, the offset sequence within P1 is [1, 2, 2, 3, 1, 0], corresponding to index 16. This mapping facilitates subsequent analysis. The phase number also records its start and end times; for example, phase P2 is labeled as "minute 712". After phase division, the control structure is formed, which can be further used to determine the fluctuation trend within each phase or to fit the control response pattern. The results of the example division are as follows:

[0101] P1: Offset value [1, 2, 2, 3, 1, 0], index [1~6];

[0102] P2: Offset [1, 2, 3, 4, 3], Index [7~11];

[0103] P3: Offset value[2], index

[12] ;

[0104] The results indicate that the switching time point can effectively divide the original offset sequence into multiple sub-segments, each reflecting the load changes over different time periods. Subsequently, feature analysis or control response models can be independently constructed for each stage. If the maximum offset value within a certain stage is more than 20% greater than the average value of the previous stage, that stage can be identified as a "load surge period," such as the maximum value of 4 in P2 and the average value of P1. The change ratio is:

[0105] ;

[0106] If the load is significantly higher than the set threshold, P2 is marked as a sudden load increase phase.

[0107] The advantage of this formula is that by calibrating and mapping the switching time points, a time segmentation structure that can be directly controlled is formed, making load response scheduling more continuous and structured.

[0108] Specifically, such as Figure 4 As shown, the specific steps of S3 are as follows:

[0109] S301: Based on the control log text recorded during the cooling load control phase, the TF-IDF algorithm is used to calculate the weight value of terms in the control statements, extract terms including chiller unit equipment names and control actions, count the number of times equipment and action combinations co-occur in the text, and obtain the equipment action co-occurrence frequency value.

[0110] The equipment action co-occurrence frequency value represents the frequency at which a certain chiller unit equipment and a certain control action appear together in the control log, and is used to measure the probability of the combination occurring.

[0111] Based on the control log text recorded during the cold load control phase, text segmentation of all control logs is required before processing. A line-by-line traversal approach is used to extract the subject, predicate, and object structure of each log entry, and sentences containing common control action verbs such as "start," "stop," "adjust," and "switch" are selected. These initially screened sentences are then compiled into a corpus for subsequent calculations. Subsequently, each sentence in the corpus is vectorized using a bag-of-words model. Without calling model functions, the frequency of word occurrences is statistically analyzed by scanning each word. The proportion of terms appearing in the log entries Next, the TF-IDF weight value of each term is calculated as follows:

[0112] ;

[0113] in, For terms In the Frequency of occurrence in the statement The number of log entries containing this term. To control the total number of log entries, for example: in 500 log entries, the word "start" appears twice in a certain statement ( The word appeared in 100 log entries. Then its weight is:

[0114] ;

[0115] After obtaining the weight value of each term, the equipment names of the chiller units (such as "Chiller No. 1", "Main Unit No. 3", "Standby Unit") and their corresponding control actions (such as "Start", "Shut Down", "Standby") are identified based on the keyword dictionary. Then, a one-to-one combination relationship is established, forming combinations such as "Chiller No. 1 - Start", "Main Unit No. 3 - Shut Down". The original text is then traversed, and each text is checked to see if the equipment name and action word appear simultaneously. If they do, the co-occurrence count of the corresponding combination is incremented by 1. The co-occurrence frequency value of the combination is obtained by calculating the ratio of the co-occurrence count of the combination to the total number of sentences in the corpus. The co-occurrence frequency is defined as:

[0116] ;

[0117] in Indicates the first The device and the first The number of times a combination of actions co-occurs. This represents the total number of sentences in the corpus. For example, if "Cold Machine No. 1 - Start-up" appears 56 times and the corpus contains 500 sentences, then the frequency value is... This method is used to extract the co-occurrence frequency value of the device-action combination.

[0118] S302: Based on the co-occurrence frequency value of equipment actions, and using the chiller unit action response threshold as a benchmark, determine whether the frequency value of each group of equipment and actions meets the response standard, filter out combinations that do not meet the benchmark, organize the combinations that meet the conditions according to equipment categories, and obtain the response combination frequency set;

[0119] The chiller unit action response threshold is the minimum frequency standard for judging whether the combination of equipment and action has responsive significance; combinations below this value will be screened out.

[0120] Based on the acquired co-occurrence frequency values ​​of equipment actions, it is necessary to determine whether the frequency value reaches the chiller unit's action response threshold. First, the threshold is set to a fixed baseline value. The equipment control frequency is set to 0.05, meaning that combinations below 5% are considered occasional occurrences and therefore do not fall within the control response range. The judgment process involves calculating the frequency values ​​of each combination... With response threshold Compare them one by one, if If the frequency value is less than 0.05, it is considered to meet the response criteria. If it does not meet the criteria, it is removed and will not be processed further. For example, the frequency value of the "Host No. 3 - Standby" combination is 0.036, which is lower than the threshold of 0.05 and is therefore removed. The "Cold Machine No. 1 - Start" combination is 0.112 and is retained. The retained combination items are aggregated and classified according to the device number. Finally, a set of response combination frequencies is constructed at the device level. The structure of multiple items in the set is in the form of triples: [Device No. 1, Action Content, Frequency Value]. For example: ["Cold Machine No. 1", "Start", 0.112], ["Cold Machine No. 2", "Off", 0.071], etc. During the sorting process, the items need to be sorted from smallest to largest according to the number. The final result is a set of combination items with a clear structure and clear device classification.

[0121] Table 3: Equipment-Action Combination Frequency Table

[0122]

[0123] As shown in Table 3, the table lists some of the device-action combinations that have passed the frequency threshold screening. All of them are valid combinations that meet the response criteria. The data in the table has been sorted in ascending order by device number to facilitate subsequent instruction generation and invocation.

[0124] S303: Call the device-action combination in the response combination frequency set, organize the combination items according to the chiller unit number order, set the standard format according to the action description and perform structural coding, integrate the action instructions under the device, and generate the chiller unit control instruction set;

[0125] The system retrieves combination items from the response combination frequency set, iterates through the combination groups corresponding to each device number, and extracts combination items in ascending order of the device number. First, the "device number" is used as the primary key field of the current processing object. During the iteration, this field is used as the classification basis for actions. Combinations corresponding to the same device number are sorted by action content. For example, under "Refrigerator No. 1," there are multiple actions such as "Start," "Shut Down," and "Adjust." Structurally, the action text is uniformly formatted as "Device Number + Action Type," such as "Refrigerator No. 1_Start," "Refrigerator No. 1_Shut Down." Then, a standard action format is set according to the action semantics, converting natural language verbs into... Standard control commands, such as "Start" in the unified format "CMD_START", "Stop" in the unified format "CMD_STOP", and "Adjust" in the format "CMD_ADJUST", are encoded in the same way for each device's action commands. Then, all action commands for that device are collected in array form, ultimately forming a structured control set such as "Chiller No. 1: {CMD_START, CMD_STOP}", "Chiller No. 2: {CMD_STOP}". The action command sets of all devices are then integrated according to their original order to form the chiller unit control command set, which is used for subsequent direct calls.

[0126] The integration process of this paragraph relies on standard semantic format recognition and structural field encoding to ensure semantic consistency and facilitate subsequent recognition and retrieval.

[0127] Specifically, such as Figure 5 As shown, the specific steps of S4 are as follows:

[0128] S401: Call the chiller unit control instruction set, based on the syntactic dependency structure tree in the scheduling request, parse the subject-verb-object relationship and modifiers, identify the parameter phrases associated with the chiller load, remove negative terms, sort the parameters according to syntactic weights, and generate the load target parameter sequence value;

[0129] When invoking the chiller unit control command set, the first step is to extract the sentence structure content included in the scheduling request. By constructing a dependency relationship structure tree based on natural language, the subject, predicate, and object in the request are syntactically identified. At the semantic level, the intention to adjust the "chiller load" is identified. For example, in the scheduling request statement "adjust the chiller load to 80% and prioritize reducing the outlet water temperature," "chiller load" is identified as the subject, "adjust" as the predicate, and "80%" as the object. "Prioritize reducing" is detected as a modifier. The modifier "reduce outlet water temperature" is further refined to extract the relevant parameter "outlet water temperature." Phrases expressing negative meanings, such as "no adjustment" and "no response required," are removed. This filtering process is then used to further refine the process. Logically, non-controllable parameters are avoided. A candidate parameter set is constructed from the extracted parameters. Each parameter in the set is weighted according to its syntactic path length. The weight is set to an integer from 1 to 5 based on the syntactic connection distance, with shorter paths having higher weights. For example, the path distance "chiller load → adjustment → 80%" is 1, and the weight is 5. The path length "chiller load → adjustment → priority → reduction → outlet water temperature" is 4, and the weight is 2. All parameters and their corresponding semantic weights are combined into an initial parameter list. The parameters in this list are then sorted in descending order by weight, resulting in a sequence of "chiller load (weight 5), outlet water temperature (weight 2)". This sequence forms the target parameter sequence value array for chiller load. Of these, 80% represents the user-defined target load value, and 7.2℃ is the set temperature value after priority reduction. This set value is determined based on the current outlet water temperature of 9.5℃ combined with the target reduction rate, resulting in a target temperature reduction of 2.3℃. Meanwhile, non-control-related modifiers in the scheduling text, such as semantically ambiguous words like "as much as possible" or "possibly," are removed. The final generated sequence of target parameters is arranged by priority as follows:

[0130] Table 4: Target Parameter Sequence Table for Chiller Load

[0131]

[0132] As shown in Table 4, the initial load is 65%, the requested target is to increase it to 80%, and the current outlet water temperature is 9.5℃. Based on the semantic "prioritize reduction" instruction, it is deduced that it needs to be adjusted to 7.2℃. This value is set within the maximum adjustable reduction of 3℃ in the temperature regulation model. Phrases with a path length greater than 5 or containing negation or conditional judgment in the grammatical analysis are removed, such as clauses like "adjust if necessary" to ensure that all extracted target parameters are valid control items. Finally, the target parameter sequence value array is output and enters the subsequent matching stage.

[0133] S402: Based on the load target parameter sequence value, for the equipment-action combination configuration, compare the adaptability of the sequence parameters with the controller action items, detect the compatibility of the response of the feedback channel with the execution command, eliminate conflicting commands, and obtain the parameter and action coordination matching rate value.

[0134] For each target load parameter sequence value, the system matches the currently executable actions of the chiller unit in the equipment-action combination configuration table. Actions available to the chiller unit's equipment controller include "start chiller," "adjust load," "adjust supply and return water temperature difference," and "switch operating modes." For each target parameter, the system retrieves the corresponding or mapped action from the controller's action list. Taking "chiller load: 80%" as an example, it identifies its compatibility with the "adjust load" action. Then, based on the current status parameters in the controller's feedback channel, it checks whether the feedback value and the selected action are in an executable state. For example, if the current operating load is already... If the execution rate reaches 75%, then 80% of the instructions will execute without conflict. If a conflict occurs, the instruction will be removed from the action set. This process is achieved by comparing whether the operation ranges of the parameter item and the equipment action item intersect. If the feedback value is at the execution boundary, for example, if the action requires adjusting the chiller outlet water temperature to 7.2℃, but the feedback shows that the chilled water pump in the current water temperature adjustment channel is already at full load, then the instruction item is determined to be an incompatible action item and is removed. In the compatibility judgment, a compatibility test is performed once for each parameter-action combination and the score is accumulated. A value of 1 is assigned for successful adaptation and 0 for failure. Finally, the parameter and action coordination matching rate is calculated using the following formula:

[0135] ;

[0136] in, To coordinate the matching rate, Indicates the first Whether the parameter is successfully adapted (1 for success, 0 for failure). Given the total number of parameter items, in the current embodiment, if the chiller load and the adjustment action are successfully matched, but the outlet water temperature action fails due to feedback limitation, then:

[0137] ;

[0138] This value indicates that the coordination between the current target parameters and the executable actions is 50%. Incompatible action items and their associated parameters need to be eliminated or adjusted and reconstructed, and the load adjustment instructions are ultimately retained.

[0139] S403: Call the parameter and action coordination matching rate value, and perform interval logic judgment on the distribution value of the parameter item in the target action range according to the upper and lower limit thresholds of the chiller unit operation. Eliminate parameters in non-controlled sections, retain the instruction parameter combination that meets the safety constraints, and generate the chiller load target parameter chain.

[0140] The coordination matching rate value is set to 0.5. Further analysis is performed on the remaining control parameters to determine if they fall within the chiller unit's permissible operating range. Considering the chiller unit's upper and lower limit thresholds, the current rated load upper limit is 100%, and the lower limit is 30%. The current target load value is 80%, falling within this range, and is therefore retained. Logical judgment is then performed on the distribution values ​​of the retained parameters within their operating range. The outlet water temperature parameter, which cannot be controlled due to feedback conflicts, is removed. The chiller load range determination is performed as follows: the target value of 80% is greater than the lower limit of 30% and less than the upper limit of 100%, meeting the requirements. According to the interval logic judgment rules, the parameter item is retained and constructed into the chiller load target parameter chain to form the final instruction parameter combination. Only the controllable instruction item "chiller load: 80%" is retained as the target value sequence for subsequent instruction execution. If the target parameter chain needs to be reconstructed, the currently incompatible items need to be logically replaced. For example, if the outlet water temperature control is unattainable, it is determined whether there are alternative items such as "return water temperature" or "supply and return water temperature difference" that are attainable parameters. The controllable item combination is reconstructed by adjusting the setting range to ensure that the final parameter chain conforms to the equipment operation constraints and feedback channel execution capabilities.

[0141] Specifically, such as Figure 6 As shown, the specific steps of S5 are as follows:

[0142] S501: Based on the load parameter item ranked first in the target parameter chain of the chiller load, retrieve the equipment-action combination item in the chiller unit control command set, screen the chilled pump as the target equipment for frequency regulation, extract its frequency range boundary and parameter item value range, establish a linear correspondence function and calculate the mapping value, and obtain the frequency response ratio value.

[0143] Based on the load parameter item ranked first in the chiller load target parameter chain, that is, the parameter with the largest value among the load data monitored by the system in the current time period is selected as the target load item. Assuming that this item is the chilled water outlet heat load, with a measured value of 850kW, the equipment-action combination items are compared one by one in the control command set. The combination items involving the adjustment frequency of the chilled pump are extracted. By parsing the execution field of the action combination, it is determined whether the chilled pump has dynamic frequency adjustment capability and is screened as the target equipment. Then, the frequency range boundary information is extracted from the equipment specification parameters of the chilled pump. For example, the lower frequency limit is 30Hz and the upper frequency limit is 60Hz. This range interval [30, 60] is used as a reference for the frequency change range. The historical value interval of the target load item is further analyzed to obtain its fluctuation range in real-time operation. Let its normal fluctuation range be [600, 1000]kW. A linear mapping function is constructed by combining the load interval and the frequency interval:

[0144] ;

[0145] Substituting the current target load of 850kW into the function, the frequency mapping value is obtained as follows:

[0146] Hz;

[0147] Considering the current operating frequency of the chilled pump is 45Hz, calculate the frequency response ratio:

[0148] As shown in Table 5.

[0149] Table 5: Frequency Range and Load Mapping Parameters

[0150]

[0151] As shown in Table 5, the mapped frequency is 48.75Hz under a given target load of 850kW. A linear proportional function is constructed by combining the lower and upper limits of the frequency to obtain a response ratio of 0.625. The determination of the frequency response ratio depends on the actual slope between the two points determined by the linear function and the mapping position at the current target load, which serves as a key weighting factor in the subsequent frequency adjustment process.

[0152] S502: Call the frequency response ratio value, read the current initial frequency value according to the device frequency adjustment threshold range, combine the frequency starting point and the ratio value to construct an adjustment function, calculate the direction and amplitude of frequency change, and generate the pump speed gain value;

[0153] Call frequency response ratio Then, first read the current initial frequency value of the chilled pump. Hz, combined with the frequency adjustment threshold range [30, 60]Hz, it is determined that the current value is in the middle range, and a linear adjustment scheme is executed. The frequency adjustment function is constructed by combining the adjustment starting point of 30Hz and the proportional value, and is set as:

[0154] ;

[0155] However, since the frequency exceeds the maximum frequency and a cropping limit is required, the frequency is adjusted as follows:

[0156] Hz;

[0157] Obtain raw frequency samples Hz, calculate the average frequency:

[0158] ;

[0159] The variance is calculated as follows:

[0160] ;

[0161] The standard deviation is Hz, current load value is 850kW, historical window load average kW, frequency-load sensitivity coefficient set to Hz / kW, calculate the pump speed gain using the formula:

[0162] ;

[0163] in, This represents the pump speed gain value and is a dimensionless parameter. This represents the current initial frequency value, in Hertz. This represents the proportional value of the frequency response, and is a dimensionless parameter. This represents the upper limit of the frequency range of the refrigeration pump, in Hertz (Hz). This represents the lower limit of the frequency range of the refrigeration pump, in Hertz. Representing the A sample of historical frequency values, in Hertz. The arithmetic mean of a sample of historical frequency values, expressed in Hertz. Represents the number of historical frequency samples, without units. This represents the frequency adjustment reference value, measured in Hertz (Hz). Represents the frequency-load sensitivity coefficient, measured in Hertz per kilowatt. This represents the value of the load parameter item that ranks first in the current chiller load target parameter chain, in kilowatts. This represents the arithmetic mean of the chiller load parameters within a certain time window, expressed in kilowatts.

[0164] Substituting the above parameters into the calculation, we get:

[0165] Numerator: ;

[0166] Denominator term: ;

[0167] Final pump speed gain value:

[0168] ;

[0169] The results indicate that the current load adjustment causes the overall frequency response to be biased upward, and the pump speed gain factor is slightly higher than 1, which means that the control action has a positive amplification trend on the frequency increase of the chilled pump, serving as the gain basis for subsequent commands.

[0170] S503: Based on the pump speed gain value and the field structure constraint rules in the chilled pump control command template, integrate the control cycle timestamp and device identifier, encapsulate the frequency change value to construct a standard data frame, write it into the device execution interface data channel, and generate an optimized execution command.

[0171] Based on the calculation of pump speed gain value The system invokes the preset field rules in the chilled pump control command template, setting the current calculation timestamp to "2025-07-23-14:25:00" and the device identifier to "CHWP-02". It then invokes the device frequency control interface template, setting the control field structure to "Device ID + Timestamp + Gain Value + Control Period + Current Frequency". The integrated data frame content is [CHWP-02, 2025-07-23-14:25:00, 1.008, 300s, 60Hz], which is written to the chilled pump communication interface data channel. The interface protocol is set to MODBUS-TCP. The interface performs standard command verification. After confirming successful field verification, the encapsulated command data packet is sent to the controller. Finally, the frequency value is applied to the frequency setting curve according to the gain factor, setting the target control frequency to [value missing]. Hz, the frequency of the device is dynamically increased by the frequency driver.

[0172] 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 variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included 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 chiller load optimization algorithm for central air conditioning chiller units based on a large language model, characterized in that, Includes the following steps: S1: Semantically analyze the natural language scheduling request text using a word vector model, extract the set of terms describing the operating status of the chiller unit, and perform cooling load intensity offset calculation on adjacent terms based on word vector spatial distance calculation to generate a cooling load intensity offset sequence. S2: Call the cooling load intensity offset sequence and iterate through it using a sliding window mechanism. When the absolute value change rate of three consecutive offsets exceeds the chiller start-up and shutdown threshold, mark it as a load switching point. Segment the text according to the switching point position to obtain the cooling load control stage. S3: Invoke the cooling load control stage, based on the TF-IDF algorithm, count the co-occurrence frequency of chiller unit equipment names and control actions, filter equipment-action combinations with frequency values ​​greater than the chiller unit action response threshold, and generate a chiller unit control instruction set; S4: Invoke the chiller unit control instruction set, extract the target parameter items directly related to the chiller load in the scheduling request through dependency parsing, sort the parameters that have not been negated, combine the physical constraints of the equipment-action combination, filter the instruction items that conflict with the safe operating range of the chiller unit, and form the chiller load target parameter chain.

2. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 1, characterized in that, The cooling load intensity offset sequence includes word vector offset, load intensity change rate and timestamp sequence; the cooling load control stage includes a set of switching time points, time period division identifier and associated temperature parameter group; the chiller unit control instruction set includes instruction coding standard, operating parameter range and validity verification identifier; the chiller load target parameter chain includes parameter priority sequence, constraint relationship network and safe operation threshold dataset.

3. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model as described in claim 1, characterized in that, The specific steps of S1 include: S101: Semantic parsing of natural language scheduling request text is performed through word vector model. After extracting keywords from the text, the associated terms of chiller unit operation status are identified. The model embedding parameters are called to convert each term into a vector representation and generate a sequence of operation status term vectors. S102: Based on the word vector sequence of the running state, extract the vector combination of adjacent words, use Euclidean distance as the measure of positional offset between vectors, calculate the spatial distance between word vectors pairwise in sequence, organize the results into a linear distance sequence, and obtain the semantic distance sequence between words; S103: Call the semantic distance sequence between the terms, match the cooling load variation level corresponding to the differential distance value according to the vector change benchmark of the chiller unit's operating condition, map the position difference to the specific intensity range according to the distance sequence order, fit the numerical correspondence between the index sequence and the intensity quantity, and generate the cooling load intensity offset sequence.

4. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model as described in claim 3, characterized in that, The specific steps of S2 include: S201: Call the cooling load intensity offset sequence, set a sliding window mechanism to traverse it, extract the absolute value of the offset within the window, calculate the rate of change of two adjacent sets of offset values ​​and organize them into a sequence to generate a continuous offset rate of change sequence. S202: Based on the continuous offset change rate sequence, combined with the chiller unit start-up and shutdown threshold, determine whether multiple consecutive values ​​exceed the limit simultaneously, record the corresponding index position and map it to the original time point to obtain the cooling load switching time point sequence. S203: Based on the cooling load switching time point sequence, call the cooling load intensity offset sequence to divide it in time order, use the position in the time point sequence as the dividing boundary to extract the interval offset data and mark the segment number to establish the cooling load control stage.

5. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 4, characterized in that, The specific steps of S3 include: S301: Based on the control log text recorded during the cooling load control phase, the TF-IDF algorithm is used to calculate the weight value of the terms in the control statement, extract the terms including the name of the chiller unit equipment and the control action, count the number of times the equipment and action combination co-occurs in the text, and obtain the equipment action co-occurrence frequency value. S302: Based on the co-occurrence frequency value of the equipment actions, and using the chiller unit action response threshold as a benchmark, determine whether the frequency value of each group of equipment and actions meets the response standard, filter out combinations that do not meet the benchmark, sort out the combinations that meet the conditions according to equipment classification, and obtain the response combination frequency set. S303: Call the device-action combination in the response combination frequency set, organize the combination items according to the chiller unit number order, set the standard format according to the action description and perform structural coding, integrate the action instructions under the device, and generate the chiller unit control instruction set.

6. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 5, characterized in that, The specific steps of S4 include: S401: Invoke the chiller unit control instruction set, based on the syntactic dependency relationship structure tree in the scheduling request, parse the subject-verb-object relationship and modifiers, identify the parameter phrases associated with the chiller load, remove negative terms, sort the parameters according to syntactic weights, and generate the load target parameter sequence value; S402: Based on the load target parameter sequence value, for the device-action combination configuration, compare the adaptability of the sequence parameters with the controller action items, detect the compatibility of the response and execution instructions of the feedback channel, eliminate conflicting instructions, and obtain the parameter and action coordination matching rate value. S403: Call the parameter and action coordination matching rate value, and perform interval logic judgment on the distribution value of the parameter item in the target action range according to the upper and lower limit thresholds of the chiller unit operation. Eliminate parameters in non-control sections, retain the instruction parameter combination that meets the safety constraints, and generate the chiller load target parameter chain.

7. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 6, characterized in that, The load target parameter sequence values ​​are an ordered set of parameters formed by extracting control parameters associated with chiller load based on syntactic dependency analysis and sorting them according to semantic weights. The parameter-action coordination matching rate value is an adaptation assessment result that measures the degree of matching between parameters and equipment control actions in terms of logic, execution, and feedback. The upper and lower operating thresholds of the chiller unit are multi-dimensional operating boundary parameters of load, temperature, and pressure that limit the safe operating range of the chiller unit.

8. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 1, characterized in that, include: S5: Based on the parameter item ranked first in the chiller load target parameter chain, call the corresponding equipment-action combination in the chiller unit control instruction set, linearly map the parameter item value with the equipment action range, calculate the chilled pump frequency adjustment percentage value, and generate optimized execution instructions. The optimized execution instructions include range conversion coefficients, frequency adjustment ranges, and timing control time nodes.

9. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 8, characterized in that, The specific steps of S5 include: S501: Based on the load parameter item ranked first in the chiller load target parameter chain, retrieve the equipment-action combination item in the chiller unit control command set, screen the chilled pump as the frequency regulation target device, extract its frequency range boundary and parameter item value range, establish a linear correspondence function and calculate the mapping value, and obtain the frequency response ratio value. S502: Call the frequency response ratio value, read the current initial frequency value according to the device frequency adjustment threshold range, construct an adjustment function by combining the frequency starting point and the ratio value, calculate the direction and amplitude of frequency change, and generate a pump speed gain value; S503: Based on the pump speed gain value and the field structure constraint rules in the refrigeration pump control command template, integrate the control cycle timestamp and device identifier, encapsulate the frequency change value to construct a standard data frame, write it into the device execution interface data channel, and generate an optimized execution instruction.

10. The chiller load optimization algorithm for central air conditioning chiller units based on a large language model according to claim 9, characterized in that, The frequency response ratio represents the linear mapping ratio of the current load parameter change within the frequency range of the chilled pump; The pump speed gain value represents the actual frequency increase or decrease of the refrigeration pump calculated based on the response ratio value; The adjustment function is a mathematical model constructed based on the response ratio and the current frequency to calculate the target frequency.

Images