An interactive one-stop energy hosting capacity load forecasting and cost optimization method and device

Abstract

The application provides an interactive one-stop energy hosting load prediction and cost optimization method and device, comprising: obtaining tenant-entered process art language description text, collecting production rhythm information and sub-meter power readings reported by each tenant, extracting a keyword sequence from the process art language description text, performing vector embedding conversion to obtain process art language embedding representation; performing similarity calculation on the process art language embedding representation and a preset synonym library, merging synonymous process arts into a unified category, associating each category with production rhythm information and sub-meter power readings to obtain process timing data sets; based on the process timing data sets and the aggregated load curve under the semantic merging condition, predicting the load of the sub-meter power readings trend of each tenant in the next production cycle of the production scheduling plan to obtain a predicted aggregated load curve.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an interactive, one-stop method for load forecasting and cost optimization in energy management. Background Technology

[0002] In the field of energy management, industrial park energy outsourcing serves as a crucial means of optimizing energy consumption and controlling costs across multiple tenants, possessing undeniable value. Particularly in industrial parks, the comprehensive analysis and resource allocation of multi-tenant electricity usage can effectively reduce overall energy consumption and operating costs. Research and practice in this area directly impact energy efficiency and the achievement of green and low-carbon goals, significantly contributing to sustainable industrial development. However, existing methods often overlook the differences in production rhythms among tenants when dealing with multi-tenant electricity usage, failing to fully explore the potential optimization opportunities arising from these differences. Many solutions focus on static total electricity consumption statistics without deeply considering the temporal distribution characteristics of tenant production activities. This neglect leads to inefficient resource allocation, particularly in the capacity planning of shared energy storage devices, often resulting in over- or under-allocation, affecting overall economic efficiency and practicality. At a deeper level, the key challenge lies in accurately identifying and utilizing the temporal differences in tenant production rhythms—that is, the phase differences in production cycles—to optimize energy storage capacity allocation. Production cycle phase difference refers to the time misalignment of high-energy-consuming processes among different tenants. This misalignment could naturally offset peak electricity consumption, reducing the demand for energy storage. However, if similar production activities of different tenants are simply categorized into the same type of event in the analysis, this time misalignment advantage will be masked, leading the system to misjudge them as synchronous high-energy-consuming events and thus overestimating energy storage demand. For example, in an industrial park, two tenants are involved in the high-energy-consuming process of stamping. If their stamping operations are scheduled in the morning and afternoon respectively, theoretically, peak electricity consumption can be staggered, the overall electricity consumption curve will be smoother, and the energy storage demand will be reduced. However, if the analysis only focuses on the common characteristic of "stamping" and ignores the difference in timing, the system may treat both as simultaneous high-energy-consuming events, mistakenly believing that a larger energy storage capacity is needed to cope with the non-existent concentrated peak. Therefore, how to fully consider the impact of production cycle phase difference when analyzing tenant electricity consumption behavior, and avoid masking the optimization space brought about by time misalignment due to simple categorization, has become a key issue that urgently needs to be addressed in the energy storage capacity configuration of park energy management. Summary of the Invention

[0003] This invention provides an interactive, one-stop energy management load forecasting and cost optimization method, mainly comprising: The process terminology description text, production cycle information, and sub-meter power readings are obtained from the tenant. The process terminology description text is processed to generate a semantic vector expression. The semantic vector expression is then bound to the production cycle information and sub-meter power readings to form an embedded representation of the process terminology. The similarity of the embedded representations of the process terms is calculated based on a preset thesaurus, and the synonymous process terms are grouped into a unified category and associated with the production cycle information and sub-meter power readings to construct a process time series dataset. The production cycle time information in the process time series dataset is grouped, and the dispersion of cycle time within each group is evaluated to determine the semantic merging condition. Based on the process time series dataset and aggregated load curve under the semantic merging conditions, the trend of sub-meter power readings for the next production cycle is predicted, and a predicted aggregated load curve is generated. The process time series dataset is classified to identify the phase complementary peak-shaving effect during peak and valley periods. The predicted aggregate load curve is corrected based on the identification results, the peak-valley difference is extracted, and the energy storage capacity configuration scheme is calculated.

[0004] Furthermore, the step of obtaining process terminology description text, production cycle time information, and sub-meter power readings from the tenant end, and processing the process terminology description text to generate semantic vector representations, includes: The process terminology description text is collected through the tenant's interactive interface, and noun phrases and verb phrases are identified from it. The frequency of each phrase in the text is counted. The word groups are sorted in descending order according to their frequency, and word groups with a frequency higher than a preset threshold are retained to generate a keyword sequence. The character length and part-of-speech category of each keyword in the keyword sequence are obtained. The keyword sequence is transformed using a preset vector embedding algorithm, and the numerical vector of each keyword in a preset dimension is output to obtain the semantic vector expression. The production cycle information and sub-meter power readings are collected synchronously. The start and end times of each process in the production cycle information are extracted. The start and end times are matched with the sampling times of the sub-meter power readings in chronological order. Based on the matching results, the semantic vector expression is bound to the corresponding production cycle information and sub-meter power readings to generate an embedded representation of process terms.

[0005] Furthermore, the step of calculating the similarity of the embedded representations of the process terms based on a preset thesaurus, grouping synonymous process terms into a unified category and associating them with the production cycle information and sub-meter power readings, includes: Obtain a preset thesaurus, which stores multiple sets of synonymous process terms and their corresponding standard category labels and standard vector expressions; For each semantic vector expression in the embedded representation of the process term, a similarity calculation algorithm is used to calculate the similarity value between it and the standard vector expression; If the similarity value is higher than the preset merging threshold, the corresponding semantic vector expression will be assigned to the corresponding standard category label to generate a category merging label; If the similarity value is lower than the preset merging threshold, the original term is retained as an independent category, and a unified category mapping table is generated. Based on the unified category mapping table, the corresponding production cycle information and sub-meter power readings are extracted from the process term embedding representation, and the labels are grouped and aggregated by category to establish an association index between each category and the time series data, thereby generating a process time series dataset.

[0006] Furthermore, the step of grouping the production cycle time information in the process time series dataset and evaluating the dispersion of cycle time within each group to determine semantic merging conditions includes: Extract the production cycle time information corresponding to each category merging label from the process time sequence dataset, and group the process start time according to the category merging label to generate cycle time groups; For each cycle time group, calculate the standard deviation of the start time of each process within the group relative to the mean to obtain the dispersion value; Obtain a preset phase discrete threshold value. If the discreteness value exceeds the phase discrete threshold value, set an independent process mark for the corresponding category merging label and lower the merging similarity threshold value to a preset low threshold value. Based on the independent process markers and the merging similarity thresholds corresponding to each category, the category merging labels are conditionally labeled to generate semantic merging conditions.

[0007] Furthermore, the prediction of the sub-meter power reading trend for the next production cycle based on the process time series dataset and aggregated load curve under the semantic merging conditions includes: From the process time series dataset under the semantic merging condition, extract the sub-table power readings and process start and end times according to the tenant workstation dimension, perform time-series alignment on the sub-table power readings, and generate an aligned power sequence. For the aligned power sequence, the power values ​​of each tenant are superimposed and summarized at the same timestamp to form time-series power data covering all tenants, and an aggregated load curve is generated. Obtain the process scheduling information for the next production cycle, match the process category according to the semantic merging condition, extract historical sub-meter power readings, use a preset prediction model to predict the trend of sub-meter power readings for the next production cycle, and output the predicted power sequence. Based on the predicted power sequence, the predicted power values ​​of each tenant are summarized in a time-aligned and superimposed manner to generate a predicted aggregated load curve.

[0008] Furthermore, classifying the process time series dataset to identify the phase complementary peak clipping effect during peak and valley periods, and correcting the predicted aggregation load curve based on the identification results, includes: Extract the sub-meter power readings and process start and end times of each category from the process time series dataset, and construct feature vectors for the power change characteristics of each category within the same time interval. The feature vector is classified using a preset classification algorithm, and the classification labels for each time interval are output to indicate whether there is a phase complementary peak clipping effect, thereby generating the complementary peak clipping period identification result. Based on the complementary peak-shaving period identification results, the time interval in the predicted aggregate load curve where the phase complementary peak-shaving effect exists is located, and the predicted power value of the time interval is reduced by a preset correction coefficient to generate the corrected predicted aggregate load curve. The peak and valley values ​​of power are extracted from the corrected predicted aggregate load curve, and the peak-valley difference is calculated.

[0009] Furthermore, the extraction of peak-valley difference and calculation of energy storage capacity configuration scheme includes: The peak power and valley power values ​​are extracted from the revised predicted aggregate load curve, and the difference between the two is calculated to obtain the peak-valley difference value. Obtain the unit price of electricity and the difference between peak and off-peak electricity prices for each tenant, and calculate the difference in electricity costs during peak and off-peak periods based on the peak-to-valley difference and the unit price of electricity. Obtain a preset energy storage unit capacity cost threshold. If the difference in electricity cost is lower than the product of the energy storage unit capacity cost threshold and the peak-valley difference value, then determine the energy storage capacity configuration value based on the ratio of the electricity cost difference to the energy storage unit capacity cost threshold, and generate an energy storage capacity reduction configuration scheme.

[0010] Furthermore, the step of transforming the keyword sequence using a preset vector embedding algorithm to output a numerical vector for each keyword in a preset dimension includes: Obtain the context phrases of each keyword in the keyword sequence as input data; For each keyword, a numerical representation of the keyword in a preset dimension is calculated using a preset vector embedding algorithm based on the semantic relevance of the context phrases.

[0011] Furthermore, the step of calculating the similarity of the embedded representations of the process terms based on a preset thesaurus, grouping synonymous process terms into a unified category and associating them with the production cycle information and sub-meter power readings, includes: For each semantic vector expression in the embedded representation of the process term, obtain the numerical features of each standard vector expression in the thesaurus; The cosine of the angle between the semantic vector expression and the standard vector expression is calculated using a preset similarity calculation algorithm to generate a similarity value. Based on the comparison result between the similarity value and the preset merging threshold, it is determined whether the semantic vector expression is classified into the corresponding standard category label; If it is classified into a standard category label, a corresponding category merging label is generated and associated with the production cycle information and sub-meter power reading; If a term is not classified into a standard category label, the original terminology expressed by the semantic vector is retained as an independent category, and the unified category mapping table is updated.

[0012] Furthermore, the step of grouping the production cycle time information in the process time series dataset and evaluating the dispersion of cycle time within each group to determine semantic merging conditions includes: Extract the numerical data of the start time of each process within the group, calculate the mean of the start time of each process within the group, and calculate the deviation value of the start time of each process based on the mean. Calculate the standard deviation of the start time of all processes within the group based on the deviation value, and generate a dispersion value; Based on the comparison result between the discreteness value and the preset phase discreteness threshold, it is determined whether there is a phase difference in the grouping; If phase differences exist, an independent process label is set for the category merging label corresponding to the group, and the merging similarity threshold is adjusted to generate the corresponding semantic merging conditions.

[0013] The present invention also provides an interactive one-stop energy management load forecasting and cost optimization device, the device comprising: The process term embedding module is used to obtain process term description text, production cycle information and sub-meter power readings from the tenant end, process the process term description text to generate semantic vector expression, and bind the semantic vector expression with the production cycle information and sub-meter power readings to form process term embedding representation; The process time series dataset construction module is used to calculate the similarity of the embedded representations of the process terms according to the preset thesaurus, group the synonymous process terms into a unified category and associate them with the production cycle information and sub-meter power readings to construct the process time series dataset. The semantic merging condition determination module is used to group the production cycle time information in the process time series dataset and evaluate the dispersion of cycle time within the group to determine the semantic merging condition. The load forecasting module is used to predict the trend of sub-meter power readings for the next production cycle based on the process time series dataset and aggregated load curve under the semantic merging conditions, and generate a predicted aggregated load curve. The energy storage capacity reduction configuration module is used to classify the process time series dataset to identify the phase complementary peak-shaving effect during peak and valley periods, correct the predicted aggregate load curve based on the identification results, extract the peak-valley difference, and calculate the energy storage capacity configuration scheme.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an interactive, one-stop energy management load forecasting and cost optimization method. It proposes an integrated solution to address the complex business problems of process semantic differences, load fluctuations, and electricity cost optimization in multi-tenant production scenarios. The invention receives process terminology descriptions entered by tenants, extracts keywords using natural language processing (NLP) technology, performs vector embedding, matches and categorizes them with a thesaurus, and associates them with production cycle time and sub-meter power readings to form a process time-series dataset. Based on this, the semantic merging conditions are dynamically adjusted to adapt to the degree of dispersion, generating an aggregated load curve and predicting the load trend for the next cycle. Simultaneously, a support vector machine (SVM) is used to identify peak-valley time-shaping effects, correct the prediction curve, and propose energy storage capacity reduction schemes based on cost accounting. This invention achieves closed-loop management of the entire process from semantic parsing to load forecasting to cost optimization, effectively reducing peak-valley differences, decreasing energy storage configuration costs, and improving energy management efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of an interactive one-stop energy management load forecasting and cost optimization method according to the present invention.

[0016] Figure 2 This is a schematic diagram of the process for identifying complementary peak-shaving periods in one embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the process for generating a reduced energy storage capacity configuration scheme in one embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram of the structure of an interactive one-stop energy management load forecasting and cost optimization device according to the present invention.

[0019] Figure 5 This is a schematic diagram illustrating the load prediction accuracy provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0021] like Figure 1 This embodiment of an interactive one-stop energy management load forecasting and cost optimization method may specifically include: Step S101: Obtain the process term description text entered by the tenant, collect the production cycle information and sub-meter power readings reported by each tenant, extract the keyword sequence from the process term description text, perform vector embedding transformation, and obtain the process term embedding representation.

[0022] The process terminology description text entered by each tenant is collected through the tenant-side interactive interface. Noun phrases and verb phrases are identified from this text. The frequency of each identified phrase in the text is counted, and the phrases are sorted in descending order of frequency. Phrasees with a frequency higher than a preset threshold are retained to obtain a keyword sequence. The character length and part-of-speech category of each keyword in the keyword sequence are obtained. The Word2Vec algorithm is used to perform vector embedding transformation on the keyword sequence. This algorithm takes the context phrases of each keyword as input and outputs a numerical vector of each keyword in a preset dimension to obtain the semantic vector expression of each keyword. Production cycle time information and sub-meter power readings reported by each tenant are synchronously collected from the tenant-side interactive interface. The start and end times of each process in the production cycle time information are extracted. The start and end times are matched with the sampling times of the sub-meter power readings in chronological order. Based on the matched times, the semantic vector expression is bound to the corresponding production cycle time information and sub-meter power reading to obtain the embedded representation of the process terminology.

[0023] In energy hosting scenarios, the tenant-side interface serves as the entry point for data collection, undertaking the function of receiving process information from each tenant.

[0024] Specifically, the interface provides each tenant with a text entry area where they can fill in text describing the process terms involved in their production process. This text includes the names of each step in the tenant's production activities, the type of equipment, and the processing method.

[0025] In one possible implementation, the collected text describing process terms is categorized using part-of-speech tagging (POS). Noun phrases typically refer to equipment or material names, while verb phrases typically refer to processing actions or operational behaviors. After tagging, the frequency of each phrase in the text is counted, with the frequency reflecting its importance in the tenant's production activities. All phrases are then sorted in descending order based on their frequency, and a frequency threshold is set. Only phrases with frequencies above this threshold are retained, thus forming a keyword sequence. This keyword sequence represents the core process terms that appear frequently in the tenant's production activities.

[0026] For example, the Word2Vec algorithm is a neural network-based word vector generation method. Its core principle is to learn the vector representation of words by analyzing the contextual co-occurrence relationships between words. In this embodiment, the input of the algorithm is each keyword in the keyword sequence and its contextual phrases in the original text. Contextual phrases refer to other phrases appearing in adjacent positions before and after the keyword. The algorithm constructs a shallow neural network to map each keyword to a numerical vector of fixed dimension, which can be preset. During training, the algorithm adjusts the vector values ​​to make the distance between frequently co-occurring phrases in the original text closer in the vector space. The output is the semantic vector representation corresponding to each keyword, and each semantic vector representation consists of a set of numerical values.

[0027] It should be noted that the production cycle information records the start and end times of each process of the tenant, while the sub-meter power readings record the power values ​​of the tenant's electricity meters at each sampling time. The time matching process involves arranging the start and end times of the processes and the power sampling times in chronological order, and finding the correspondence between power sampling times falling within a certain process start and end time interval.

[0028] In one embodiment, after time-based matching is completed, the semantic vector representation is bound to the production cycle information and sub-meter power readings within the corresponding time period. The binding method involves creating a data record containing three fields: a semantic vector representation field storing the vector values ​​of keywords, a production cycle information field storing the start and end times of the process, and a sub-meter power reading field storing the power values ​​within that time period. Multiple data records are aggregated to form a process term embedding representation, which simultaneously includes the semantic features of the process terms and their corresponding time-series power consumption characteristics.

[0029] Step S102: The process term embedding representation is compared with the preset thesaurus for similarity calculation. Synonymous process terms are grouped into a unified category. Each category is then associated with production cycle information and sub-meter power readings to obtain a process time series dataset.

[0030] A pre-defined thesaurus is obtained, which stores multiple sets of synonymous process terms and their corresponding standard category labels. Each standard category label is pre-associated with a standard vector expression. For each semantic vector expression in the process term embedding representation, a cosine similarity algorithm is used to calculate the cosine value of the angle between it and each standard vector expression in the thesaurus, obtaining the similarity value between each semantic vector expression and each standard category label. If the similarity value between a semantic vector expression and a standard category label is higher than a pre-defined merging threshold, the semantic vector expression is assigned to that standard category label, generating a category merging label; if the similarity values ​​are all lower than the merging threshold, the original term of the semantic vector expression is retained as an independent category, resulting in a unified category mapping table. Based on the category merging labels in the unified category mapping table, the corresponding production cycle information and sub-meter power readings are extracted from the process term embedding representation. The process start and end times of the production cycle information and the sampled values ​​of the sub-meter power readings are grouped and aggregated according to the category merging labels, establishing an association index between each category and the time series data, resulting in a process time series dataset.

[0031] In the context of energy management, the thesaurus is a pre-established standardized lookup table of process terms.

[0032] Specifically, the thesaurus is classified according to common industrial production process types. Each category corresponds to a standard category label. Each standard category label contains multiple process terms that express the same or similar meanings. At the same time, each standard category label is pre-associated with a standard vector expression, which is calculated by taking the average of the vectors of all synonyms under that category.

[0033] In one possible implementation, the cosine similarity algorithm measures the similarity between two vectors by calculating the cosine of the angle between them. The algorithm takes as input the semantic vector representations in the embedded representation of the technological term and the standard vector representations in the thesaurus. The calculation involves dividing the inner product of the two vectors by the product of their magnitudes. The output cosine value ranges from -1 to +1; the closer the value is to +1, the more consistent the directions of the two vectors are, and the closer their semantics are. In an industrial park application, if a tenant enters the technological term "stamping," and the standard vector representation of the "stamping processing" category in the thesaurus has a cosine value of 0.92 with it, it indicates that the two are highly semantically similar.

[0034] It should be noted that the merging threshold is a pre-set similarity judgment boundary value. When the similarity value between a semantic vector expression and a standard category label is higher than the merging threshold, they are determined to belong to the same semantic category, and the semantic vector expression is assigned to that standard category label.

[0035] For example, if one tenant enters "hydraulic punching processing" and another tenant enters "stamping forming operation," and after cosine similarity calculation, the similarity values ​​of both with the standard category label "stamping processing" are higher than the merging threshold, then both are categorized into the "stamping processing" category, forming a unified category merging label. If the similarity values ​​of a term entered by a tenant with all standard category labels are lower than the merging threshold, then that term is retained as an independent category. Based on the above merging results, a unified category mapping table records the correspondence between each semantic vector expression and its belonging category. According to the category merging labels in this mapping table, the corresponding production cycle information and sub-table power readings are extracted from the embedded representation of the process terms. The start and end times of the process and the power sampling values ​​are grouped and aggregated according to the category merging labels. Multiple records under the same category are aggregated to form a time-series data subset of that category. The time-series data subsets of all categories together constitute the process time-series dataset.

[0036] Step S103: Group the production cycle information in the process time series dataset, evaluate the dispersion of each cycle time in the group, evaluate the semantic merging condition based on the dispersion, and for groups whose phase dispersion exceeds the preset threshold, lower the merging similarity threshold to retain the distinction of independent processes, and determine the semantic merging condition.

[0037] Production cycle time information corresponding to each category merging label is extracted from the process time sequence dataset. The process start times in the production cycle time information are grouped according to the category merging label. All process start times under the same category merging label constitute a cycle time group, resulting in cycle time groups for each category. For each cycle time group, the values ​​of the start times of each process within the group are extracted, and the standard deviation of all process start times within the group relative to the group mean is calculated. This standard deviation value characterizes the dispersion of each cycle time within the group, resulting in a dispersion value for each cycle time group. A preset phase dispersion threshold is obtained. The dispersion value of each cycle time group is compared with the phase dispersion threshold. If the dispersion value of a certain cycle time group exceeds the phase dispersion threshold, it is determined that the group has a phase difference. An independent process label is set for the category merging label corresponding to the group, and the merging similarity threshold value of that category is lowered to a preset low threshold value. Based on the independent process label and the merging similarity threshold value corresponding to each category, conditional annotations are performed on the merging labels of each category in the process time series dataset. Categories without independent process labels are assigned the original merging similarity threshold value, while categories with independent process labels are assigned the lower threshold value, thus obtaining the semantic merging conditions corresponding to each category.

[0038] In one possible implementation, a unified category mapping table is used to extract category merging labels. The input is the mapping table and the dataset. The k-means algorithm, where k is the number of categories, aggregates labels with a similarity higher than 0.8, and outputs an updated association index, ensuring the mapping table is continuously applied to subsequent analyses. In energy hosting scenarios, the process time-series dataset contains multiple category merging labels and their corresponding production cycle time information. Grouping the process start times in the production cycle time information according to the category merging labels is a fundamental operation for evaluating the time distribution characteristics within each category.

[0039] Specifically, under the same category's merge tag, there may be multiple production records from different tenants. Each record corresponds to a process start time, and these times constitute the cycle time grouping for that category. Standard deviation is a statistic that measures the dispersion of a set of values; a larger value indicates a more significant deviation between the values ​​within the group and the mean. In this implementation, for each cycle time group, the values ​​of the process start times within the group are extracted, and the standard deviation of these times relative to the group mean is calculated.

[0040] For example, if three tenants in an industrial park are all involved in the "injection molding" process, and their process starts at 8:00 AM, 9:30 AM, and 2:00 PM respectively, then the mean of this group is approximately 11:10 AM. The deviations of each time from the mean are large, and the calculated standard deviation is high, indicating that there is a significant time misalignment in the production cycle of each tenant in this group.

[0041] It should be noted that the phase dispersion threshold is a pre-defined judgment boundary used to distinguish whether the clock timing groups belong to a concentrated or dispersed distribution. When the dispersion value of a certain group exceeds the threshold, it is determined that the group has a phase difference, which means that the high-energy-consuming processes of each tenant under this category do not occur simultaneously, but rather exhibit staggered distribution characteristics on the time axis.

[0042] For example, in the actual operation of an industrial park, phase differences have significant business implications. If the stamping processes of two tenants under the "stamping processing" category are scheduled for morning and afternoon shifts respectively, their peak electricity consumption will naturally be staggered, resulting in a relatively flat overall electricity consumption curve for the park. Simply categorizing them together for uniform processing would mask the peak-shaving potential brought about by this time misalignment. By setting independent process tags, the distinction between processes with phase differences can be preserved, preventing them from being treated as synchronous electricity loads in subsequent processing.

[0043] In one embodiment, the setting of independent process tags is based on the comparison result of the dispersion value and the phase dispersion threshold value. When the dispersion value exceeds the threshold value, an independent process tag is added to the category merging label corresponding to the group. This tag indicates that there is a significant difference in production cycle phase within the category, and a stricter distinction standard should be adopted in the subsequent merging judgment. Furthermore, lowering the merging similarity threshold value is a specific means of achieving independent process distinction. The original merging similarity threshold value is suitable for categories with concentrated distribution of cycle times. For these categories, a higher similarity threshold value is sufficient to complete reasonable merging. However, for categories with phase differences, using the original threshold value may lead to the incorrect merging of processes with different production cycles. By lowering the merging similarity threshold value of this category to a preset low threshold value, the strictness of the merging judgment is increased. Only terms with higher similarity will be classified into the same category, thereby preserving the independent distinction of processes with different production cycles.

[0044] Specifically, the low-level threshold is a pre-set, relatively low value. Its setting principle is to minimize erroneous merging of processes with different time distributions while ensuring basic semantic similarity judgment. In industrial park applications, the specific value of the low-level threshold can be adjusted according to the typical time distribution characteristics of various processes within the park to adapt to different park production scheduling patterns. Based on the aforementioned independent process labels and the merging similarity threshold values ​​corresponding to each category, conditional annotations are applied to the merging labels of each category in the process time-series dataset. Categories without independent process labels indicate relatively concentrated production cycles, and the original merging similarity threshold value is used as the merging judgment boundary; categories with independent process labels indicate phase differences within them, and the low-level threshold value is used as the merging judgment boundary.

[0045] Understandably, semantic merging conditions are the judgment rules that each category should follow in subsequent merging processes, including the merging similarity threshold value of the category and whether it has independent process labels. By setting differentiated semantic merging conditions, the impact of production cycle phase differences can be fully considered in the semantic merging process of process terms, and processes with time misalignments can be independently identified, avoiding the potential advantage of mutual cancellation of peak electricity consumption caused by production cycle misalignment between tenants due to simple merging.

[0046] Step S104: Based on the process time series dataset and aggregated load curve under semantic merging conditions, load prediction is performed on the power reading trend of each tenant's sub-meter in the next production cycle of the production schedule to obtain the predicted aggregated load curve.

[0047] From the process time-series dataset under semantic merging conditions, the sub-meter power readings and process start and end times for each tenant are extracted according to the tenant workstation dimension. The sub-meter power readings for each tenant are time-aligned based on the process start and end times. Power readings from different tenants within the same time interval are arranged by timestamp to obtain the aligned power sequence for each tenant. For the aligned power sequence, the power values ​​of each tenant are superimposed and summarized at the same timestamp position. The power consumption of each tenant during the process operation period is accumulated to the corresponding time interval, forming time-series power data covering the power load of all tenants, resulting in an aggregated load curve. The process scheduling information for the next production cycle in the production plan is obtained. The process category of each tenant in the scheduling information is matched according to the semantic merging conditions. Historical sub-meter power readings corresponding to the scheduled process categories are extracted from the process time-series dataset. A Long Short-Term Memory (LSTM) network is used to predict the trend of sub-meter power readings for each tenant in the next production cycle. This network takes historical sub-meter power readings and process start and end times as input and outputs the predicted power sequence for each tenant. Based on the predicted power sequence of each tenant, and following the same time alignment and superposition method as the aggregated load curve, the predicted power values ​​of each tenant are summarized and superimposed in each time interval of the next production cycle to obtain the predicted aggregated load curve.

[0048] In the scenario of energy management in industrial parks, the process time-series dataset is categorized and labeled according to semantic merging conditions. Extracting the sub-meter power readings and process start and end times for each tenant from this dataset according to the tenant workstation dimension is the basic operation for constructing the aggregated load curve. The tenant workstation dimension refers to using the production workstation of an individual tenant as the data extraction unit, with each workstation corresponding to a set of independent sub-meter power reading time-series records.

[0049] In one possible implementation, timing alignment involves arranging power readings from different tenants within the same time interval according to a uniform time scale.

[0050] Specifically, the sampling frequency and sampling start time of the sub-meter power readings of each tenant may differ. The timing alignment operation determines the effective data range of each tenant based on the start and end times of the process. Within this range, the power readings are interpolated or resampled according to a uniform time step, so that the power data of different tenants have corresponding values ​​at the same timestamp position, forming the aligned power sequence of each tenant.

[0051] It should be noted that the aggregated load curve is a power time-series data reflecting the overall electricity load of the park over time. After obtaining the aligned power sequence of each tenant, the power values ​​of all tenants at each timestamp are summed up, and the summed result is the total electricity consumption of the park at that time. The summed results of all timestamps are arranged and connected in chronological order to form a complete power time-series curve, which is the aggregated load curve.

[0052] For example, in an industrial park, there are three tenants. Tenant A's stamping process operates from 8:00 AM to 10:00 AM, during which the sub-meter power reading is approximately 50 kW; Tenant B's injection molding process operates from 9:00 AM to 11:00 AM, during which the sub-meter power reading is approximately 40 kW; and Tenant C's welding process operates from 10:00 AM to 12:00 PM, during which the sub-meter power reading is approximately 35 kW. After time alignment, during the 9:00 AM to 10:00 AM time interval, Tenant A and Tenant B's processes operate simultaneously, with a combined power of 90 kW; during the 10:00 AM to 11:00 AM time interval, Tenant B and Tenant C's processes operate simultaneously, with a combined power of 75 kW. Further, the production scheduling plan records the process arrangements of each tenant in the park for future production cycles, including the planned start and end times and process categories for each process. The process scheduling information for the next production cycle is obtained from the production scheduling plan. Based on semantic merging conditions, the process categories of each tenant in the scheduling information are matched to identify the category merging label to which each scheduled process belongs. Based on the matching results, historical sub-meter power readings corresponding to the scheduled process categories are extracted from the process time series dataset. This historical data serves as a reference for prediction.

[0053] In one embodiment, the Long Short-Term Memory (LSTM) network is a recurrent neural network structure capable of remembering long-term dependencies in time-series data. The network includes three gating units: an input gate, a forget gate, and an output gate, controlling the retention and forgetting of information through a gating mechanism. In this embodiment, the LSM network takes historical sub-meter power readings and process start and end times as inputs. The historical sub-meter power readings provide the temporal variation pattern of power values, while the process start and end times provide the time boundary information of process operation. By learning the correspondence between power changes and process times in historical data, the network outputs predicted power values ​​for each tenant in each time interval of the next production cycle, forming a predicted power sequence for each tenant.

[0054] Specifically, during the training phase, the Long Short-Term Memory (LSTM) network receives sub-meter power readings and process start and end times from multiple historical production cycles, learning typical power variation patterns for different process categories across different time intervals. In the prediction phase, the network, based on the process start and end times information in the production schedule and the learned power variation patterns, infers the sub-meter power reading trends for each tenant during each process's runtime in the next production cycle. Based on the predicted power sequences for each tenant, the data is processed using the same time-series alignment and overlay method as the aggregated load curve. Within each time interval of the next production cycle, the predicted power values ​​for each tenant are accumulated and summarized at the same timestamp, forming predicted power time-series data covering the entire time interval of the next production cycle.

[0055] Understandably, the predicted aggregated load curve reflects the expected overall electricity load trend of the industrial park in the next production cycle. The peak position and magnitude of the curve indicate the concentrated periods and degree of electricity load in the park in the next cycle, while the trough position indicates the low periods of electricity load. By comparing the predicted aggregated load curve with the historical aggregated load curve, the changing characteristics of the electricity load distribution in the next production cycle relative to the historical cycle can be identified.

[0056] Step S105: Classify the process time series dataset under semantic merging conditions, identify whether the peak and valley periods in the sub-meter power readings show a phase complementary peak-shaving effect, obtain the complementary peak-shaving period identification results to correct the predicted aggregate load curve, extract its peak-valley drop, and calculate the energy storage capacity reduction configuration scheme.

[0057] like Figure 2 From the process time series dataset under semantic merging conditions, power readings and process start and end times for each category are extracted. Feature vectors are constructed for the power change characteristics of each category within the same time interval. The feature vectors contain three components: power peak time, power valley time, and peak-valley duration. The feature vectors are normalized by dividing each component by its maximum value to obtain a value in the range of 0 to 1, which is used as input to the support vector machine algorithm. The algorithm uses a linear kernel function and trains the model through 5-fold cross-validation. Using the feature vectors as input, the algorithm outputs classification labels for each time interval. The classification labels indicate whether there is a phase complementary peak clipping effect in the time interval, thus obtaining the complementary peak clipping period identification result.

[0058] like Figure 3Based on the complementary peak-shaving period identification results, the time intervals marked as having a phase complementary peak-shaving effect are located in the predicted aggregated load curve. The predicted power values ​​within these time intervals are adjusted downwards according to a preset correction coefficient of 0.8, determined by analyzing the peak reduction ratio in historical load data, resulting in the corrected predicted aggregated load curve. The peak and valley power values ​​are extracted from the corrected predicted aggregated load curve, and the difference between them is calculated to obtain the peak-valley difference value. The electricity unit price and peak-valley electricity price difference for each tenant are obtained from the power grid database. Based on the peak-valley difference value and the electricity unit price, the electricity cost difference during peak and valley periods is calculated using the formula D=(P U) E represents the energy storage unit capacity cost difference, where D is the peak-valley difference, P is the peak-valley price difference, U is the unit price of electricity, and E is the peak-valley price difference. A preset energy storage unit capacity cost threshold of 0.5 yuan per kilowatt-hour is obtained, based on the average market cost of energy storage equipment. If the energy storage unit capacity cost difference is lower than the product of the energy storage unit capacity cost threshold and the peak-valley difference, the energy storage capacity configuration value is determined according to the ratio of the energy storage unit capacity cost difference to the energy storage unit capacity cost threshold, resulting in an energy storage capacity reduction configuration scheme.

[0059] In the scenario of energy management in industrial parks, the process time-series dataset under semantic merging conditions has already completed the classification and condition labeling of process categories. Extracting the power readings and start and end times of each process for each category from this dataset forms the data foundation for peak and valley period feature identification. Based on the power change characteristics of each category within the same time interval, feature vectors are constructed for classification judgment.

[0060] Specifically, the feature vector comprises three components: power peak time, power trough time, and peak-trough duration. The power peak time refers to the point in time when the power value of this category reaches its maximum value within a certain time interval; the power trough time refers to the point in time when the power value reaches its minimum value; and the peak-trough duration refers to the time span between the trough time and the peak time. These three components together describe the power change rhythm characteristics of this category within a specific time interval.

[0061] In one possible implementation, the Support Vector Machine (SVM) algorithm is a binary classification method based on the principle of maximizing the margin. This algorithm searches for a hyperplane in the feature space that maximizes the distance from sample points of different classes to this hyperplane. In this embodiment, the SVM algorithm takes the feature vectors of each time interval as input and learns the correspondence between the feature vectors of labeled samples and the classification labels to output the classification labels for each time interval. The classification labels are divided into two categories: one category indicates that a phase complementary peak-shaving effect exists in the time interval, and the other category indicates that the effect does not exist.

[0062] It should be noted that the phase complementary peak shaving effect refers to the phenomenon that, within the same time interval, the peak and trough electricity consumption of different tenants are staggered in time, thus reducing the fluctuation range of the overall electricity load of the park relative to the sum of the fluctuation ranges of the individual electricity loads of each tenant.

[0063] For example, in an industrial park, tenant A's stamping process reaches its power peak at 9:00 AM, while tenant B's injection molding process is at its power trough at 9:00 AM. The power changes of the two complement each other in time, and the overall power of the park during that period is reduced.

[0064] For example, when the support vector machine algorithm determines that a certain time interval has a classification label indicating the presence of a phase complementary peak-shaving effect, it indicates that there is a favorable time misalignment in the production cycle of each tenant within that interval. The complementary peak-shaving period identification results record all time intervals identified as having this effect and their corresponding classification labels. Further, the predicted aggregated load curve is corrected based on the complementary peak-shaving period identification results. The time intervals in the predicted aggregated load curve identified as having a phase complementary peak-shaving effect are located, and the predicted power values ​​within these time intervals are adjusted downwards according to a preset correction coefficient. The correction coefficient is a pre-set value less than 1, and its function is to reflect the degree to which the phase complementary effect reduces the overall power peak of the park. After correction, the power values ​​of the predicted aggregated load curve during the complementary peak-shaving period are correspondingly reduced, forming the corrected predicted aggregated load curve.

[0065] In one embodiment, the peak-to-valley difference value is extracted by identifying the power peak and power trough values ​​from the corrected predicted aggregate load curve and calculating the difference between them. This difference reflects the fluctuation range of the park's electricity load in the next production cycle; a larger peak-to-valley difference value indicates a greater power gap between the peak and trough of the park's electricity load. Understandably, electricity cost accounting involves the unit price of electricity for each tenant and the peak-valley electricity price difference. The unit price of electricity refers to the price a tenant pays for electricity during normal times; the peak-valley electricity price difference refers to the difference between the electricity price during peak and off-peak hours. Based on the product of the peak-valley difference and the unit price of electricity, combined with the peak-valley electricity price difference, the peak-valley electricity cost difference is calculated. This difference represents the amount of electricity cost saved through peak shaving and valley filling operations. The specific formula can be expressed as follows: , This represents the difference in electricity costs during peak and off-peak hours. This represents the peak-to-valley difference in value, where 'c' represents the unit price of electricity. This represents the peak-valley electricity price difference. The formula calculates the electricity cost savings by multiplying the peak-valley difference by the unit price and then combining this product with the electricity price difference. The determination of the energy storage capacity reduction configuration scheme is based on a comparison between the electricity cost difference and the energy storage unit capacity cost threshold. The energy storage unit capacity cost threshold is a pre-set value representing the construction and operation cost per unit capacity of the energy storage equipment. If the electricity cost difference is lower than the product of the energy storage unit capacity cost threshold and the peak-valley difference, it indicates that the economic benefits of peak shaving and valley filling operations through energy storage equipment are low. In this case, the energy storage capacity configuration value is determined according to the ratio of the electricity cost difference to the energy storage unit capacity cost threshold. This configuration value is lower than the energy storage capacity configured directly based on the peak-valley difference value, thereby achieving a reduction in energy storage capacity configuration and avoiding resource waste caused by over-configuration of energy storage equipment.

[0066] like Figure 4 The present invention also provides an interactive one-stop energy management load forecasting and cost optimization device, the device comprising: The process term embedding module is used to obtain process term description text, production cycle information and sub-meter power readings from the tenant end, process the process term description text to generate semantic vector expression, and bind the semantic vector expression with the production cycle information and sub-meter power readings to form process term embedding representation; The process time series dataset construction module is used to calculate the similarity of the embedded representations of the process terms according to the preset thesaurus, group the synonymous process terms into a unified category and associate them with the production cycle information and sub-meter power readings to construct the process time series dataset. The semantic merging condition determination module is used to group the production cycle time information in the process time series dataset and evaluate the dispersion of cycle time within the group to determine the semantic merging condition. The load forecasting module is used to predict the trend of sub-meter power readings for the next production cycle based on the process time series dataset and aggregated load curve under the semantic merging conditions, and generate a predicted aggregated load curve. The energy storage capacity reduction configuration module is used to classify the process time series dataset to identify the phase complementary peak-shaving effect during peak and valley periods, correct the predicted aggregate load curve based on the identification results, extract the peak-valley difference, and calculate the energy storage capacity configuration scheme.

[0067] like Figure 5 , Figure 5This diagram illustrates the comparison of load prediction accuracy provided by an embodiment of the present invention. The horizontal axis represents the next production cycle time in hours; the vertical axis represents the aggregated load power of the industrial park in kilowatts. In the diagram, the thick solid curve A represents the prediction result of the traditional fixed semantic merging method, the long dashed curve B represents the prediction result of the present invention with dynamically adjusted semantic merging conditions, and the dotted-dash curve C represents the actual aggregated load curve of the industrial park. The traditional method simply merges semantically similar processes such as "stamping" and "forming" that are executed at off-peak times, easily misjudging them as synchronous high-energy-consuming processes, resulting in prediction results significantly higher than actual measurements at peak values ​​P1 and P2. The method of the present invention assesses the dispersion of production cycle time, sets independent markers for processes with phase differences, and lowers the merging similarity threshold, preserving the time misalignment characteristics during the prediction stage. Therefore, curve B is closer to the actual aggregated load curve C in each time interval. The figure shows that the production cycle phase difference is an objective temporal characteristic. By quantifying it through the degree of dispersion assessment as the basis for adjusting the merging conditions, the prediction model can more realistically reflect the asynchronous superposition effect of multi-tenant electricity consumption behavior on the time axis, which is in line with the principle of linear additivity of load superposition in the power system.

[0068] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. An interactive, one-stop energy management load forecasting and cost optimization method, characterized in that, include: The process terminology description text, production cycle information, and sub-meter power readings are obtained from the tenant. The process terminology description text is processed to generate a semantic vector expression. The semantic vector expression is then bound to the production cycle information and sub-meter power readings to form an embedded representation of the process terminology. The process term embedding representations are similar to each other based on a preset thesaurus. Synonymous process terms are then grouped into a unified category and associated with the production cycle information and sub-meter power readings to construct a process time-series dataset. This process includes: acquiring a preset thesaurus, which stores multiple sets of synonymous process terms and their corresponding standard category labels and standard vector expressions; calculating the similarity value between each semantic vector expression in the process term embedding representation and the standard vector expression using a similarity calculation algorithm; and if the similarity value is higher than a preset merging threshold, the corresponding semantic vector expression is grouped under the corresponding standard category label, generating a category merging label. The production cycle time information in the process time series dataset is grouped, and the dispersion of cycle time within each group is evaluated to determine semantic merging conditions. This includes: extracting production cycle time information corresponding to each category merging label from the process time series dataset; grouping the process start time according to the category merging label to generate cycle time groups; for each cycle time group, calculating the standard deviation of the start time of each process within the group relative to the mean to obtain a dispersion value; obtaining a preset phase dispersion threshold; if the dispersion value exceeds the phase dispersion threshold, setting an independent process label for the corresponding category merging label and lowering the merging threshold to a preset low threshold; and conditionally labeling the category merging label according to the independent process label and the merging threshold corresponding to each category to generate semantic merging conditions. The semantic merging conditions include two pieces of information: the merging threshold for each category and whether there is an independent process label. Based on the process time series dataset and aggregated load curve under the semantic merging conditions, the trend of sub-meter power readings for the next production cycle is predicted, and a predicted aggregated load curve is generated. This includes: extracting sub-meter power readings and process start and end times from the process time series dataset under the semantic merging conditions according to the tenant workstation dimension; performing time-series alignment on the sub-meter power readings to generate an aligned power sequence; and superimposing and summarizing the power values ​​of each tenant at the same timestamp position for the aligned power sequence to form time-series power data covering all tenants, thereby generating an aggregated load curve. The process time series dataset is classified to identify the phase complementary peak-shaving effect during peak and valley periods. The predicted aggregate load curve is corrected based on the identification results, the peak-valley difference is extracted, and the energy storage capacity configuration scheme is calculated.

2. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The process involves obtaining process terminology description text, production cycle time information, and sub-meter power readings from the tenant end, and processing the process terminology description text to generate a semantic vector representation, including: The process terminology description text is collected through the tenant's interactive interface, and noun phrases and verb phrases are identified from it. The frequency of each phrase in the text is counted. The word groups are sorted in descending order according to their frequency, and word groups with a frequency higher than a preset threshold are retained to generate a keyword sequence. The character length and part-of-speech category of each keyword in the keyword sequence are obtained. The keyword sequence is transformed using a preset vector embedding algorithm, and the numerical vector of each keyword in a preset dimension is output to obtain the semantic vector expression. The production cycle information and sub-meter power readings are collected synchronously. The start and end times of each process in the production cycle information are extracted. The start and end times are matched with the sampling times of the sub-meter power readings in chronological order. Based on the matching results, the semantic vector expression is bound to the corresponding production cycle information and sub-meter power readings to generate an embedded representation of process terms.

3. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The step of calculating the similarity of the embedded representations of the process terms according to a preset thesaurus, grouping synonymous process terms into a unified category and associating them with the production cycle information and sub-meter power readings, further includes: If the similarity value is lower than the preset merging threshold, the original term is retained as an independent category, and a unified category mapping table is generated. Based on the unified category mapping table, the corresponding production cycle information and sub-meter power readings are extracted from the process term embedding representation, and the labels are grouped and aggregated by category to establish an association index between each category and the time series data, thereby generating a process time series dataset.

4. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The method of predicting the trend of sub-meter power readings for the next production cycle based on the process time series dataset and aggregated load curve under the semantic merging conditions also includes: Obtain the process scheduling information for the next production cycle, match the process category according to the semantic merging condition, extract historical sub-meter power readings, use a preset prediction model to predict the trend of sub-meter power readings for the next production cycle, and output the predicted power sequence. Based on the predicted power sequence, the predicted power values ​​of each tenant are summarized in a time-aligned and superimposed manner to generate a predicted aggregated load curve.

5. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The process of classifying the time-series dataset to identify the phase complementary peak-shaving effect during peak and valley periods, and correcting the predicted aggregation load curve based on the identification results, includes: Extract the sub-meter power readings and process start and end times of each category from the process time series dataset, and construct feature vectors for the power change characteristics of each category within the same time interval. The feature vector is classified using a preset classification algorithm, and the classification labels for each time interval are output to indicate whether there is a phase complementary peak clipping effect, thereby generating the complementary peak clipping period identification result. Based on the complementary peak-shaving period identification results, the time interval in the predicted aggregate load curve where the phase complementary peak-shaving effect exists is located, and the predicted power value of the time interval is reduced by a preset correction coefficient to generate the corrected predicted aggregate load curve. The peak and valley values ​​of power are extracted from the corrected predicted aggregate load curve, and the peak-valley difference is calculated.

6. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The scheme for extracting peak-valley differences and calculating energy storage capacity configuration includes: The peak power and valley power values ​​are extracted from the revised predicted aggregate load curve, and the difference between the two is calculated to obtain the peak-valley difference value. Obtain the unit price of electricity and the difference between peak and off-peak electricity prices for each tenant, and calculate the difference in electricity costs during peak and off-peak periods based on the peak-to-valley difference and the unit price of electricity. Obtain a preset energy storage unit capacity cost threshold. If the difference in electricity cost is lower than the product of the energy storage unit capacity cost threshold and the peak-valley difference value, then determine the energy storage capacity configuration value based on the ratio of the electricity cost difference to the energy storage unit capacity cost threshold, and generate an energy storage capacity reduction configuration scheme.

7. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 2, characterized in that, The step of transforming the keyword sequence using a preset vector embedding algorithm to output a numerical vector for each keyword in a preset dimension includes: Obtain the context phrases of each keyword in the keyword sequence as input data; For each keyword, a numerical representation of the keyword in a preset dimension is calculated using a preset vector embedding algorithm based on the semantic relevance of the context phrases.

8. The interactive one-stop energy management load forecasting and cost optimization method as described in claim 1, characterized in that, The step of calculating the similarity of the embedded representations of the process terms according to a preset thesaurus, grouping synonymous process terms into a unified category and associating them with the production cycle information and sub-meter power readings includes: For each semantic vector expression in the embedded representation of the process term, obtain the numerical features of each standard vector expression in the thesaurus; The cosine of the angle between the semantic vector expression and the standard vector expression is calculated using a preset similarity calculation algorithm to generate a similarity value. Based on the comparison result between the similarity value and the preset merging threshold, it is determined whether the semantic vector expression is classified into the corresponding standard category label; If it is classified into a standard category label, a corresponding category merging label is generated and associated with the production cycle information and sub-meter power reading; If a term is not classified into a standard category label, the original terminology expressed by the semantic vector is retained as an independent category, and the unified category mapping table is updated.

9. An interactive one-stop energy management load forecasting and cost optimization device, characterized in that, The device includes: The process term embedding module is used to obtain process term description text, production cycle information and sub-meter power readings from the tenant end, process the process term description text to generate semantic vector expression, and bind the semantic vector expression with the production cycle information and sub-meter power readings to form process term embedding representation; The process time-series dataset construction module is used to calculate the similarity of the embedded representations of process terms based on a preset thesaurus, group synonymous process terms into a unified category and associate them with the production cycle information and sub-meter power readings to construct a process time-series dataset. This includes: acquiring a preset thesaurus, which stores multiple sets of synonymous process terms and their corresponding standard category labels and standard vector expressions; calculating the similarity value between each semantic vector expression in the embedded representation of the process terms and the standard vector expression using a similarity calculation algorithm; and if the similarity value is higher than a preset merging threshold, grouping the corresponding semantic vector expression into the corresponding standard category label to generate a category merging label. The semantic merging condition determination module is used to group the production cycle time information in the process time series dataset, evaluate the dispersion of cycle time within each group, and determine the semantic merging condition. This includes: extracting the production cycle time information corresponding to each category merging label from the process time series dataset; grouping the process start times according to the category merging labels to generate cycle time groups; for each cycle time group, calculating the standard deviation of the start times of each process within the group relative to the mean to obtain a dispersion value; obtaining a preset phase dispersion threshold; if the dispersion value exceeds the phase dispersion threshold, setting an independent process label for the corresponding category merging label and lowering the merging threshold to a preset low threshold; and conditionally labeling the category merging labels according to the independent process label and the merging threshold corresponding to each category to generate semantic merging conditions. The semantic merging conditions include two pieces of information: the merging threshold for each category and whether there is an independent process label. The load forecasting module is used to predict the trend of sub-meter power readings for the next production cycle based on the process time-series dataset and aggregated load curve under the semantic merging conditions, and generate a predicted aggregated load curve. This includes: extracting sub-meter power readings and process start and end times from the process time-series dataset under the semantic merging conditions according to the tenant workstation dimension; performing time-series alignment on the sub-meter power readings to generate an aligned power sequence; and superimposing and summarizing the power values ​​of each tenant at the same timestamp position for the aligned power sequence to form time-series power data covering all tenants, thereby generating an aggregated load curve. The energy storage capacity reduction configuration module is used to classify the process time series dataset to identify the phase complementary peak-shaving effect during peak and valley periods, correct the predicted aggregate load curve based on the identification results, extract the peak-valley difference, and calculate the energy storage capacity configuration scheme.

Images