Mechanism-based enhanced and interpretable double-layer load forecasting method, device and equipment

Abstract

This application provides a two-tiered load forecasting method, apparatus, and equipment based on mechanism enhancement and interpretability. The method involves constructing physical functional features from relevant characteristic data of an integrated energy system to obtain physical mechanism enhancement features corresponding to the integrated energy system; extracting temporal features from historical operating loads to obtain deep temporal embedding features corresponding to the integrated energy system; these deep temporal embedding features characterize the temporal dependence of the integrated energy system's operating load within historical time windows; fusing the physical mechanism enhancement features and the deep temporal embedding features to obtain fused features; and performing forecasting based on these fused features to obtain the predicted load of the integrated energy system and corresponding interpretable information. The interpretable information characterizes the importance of each feature in the physical mechanism enhancement features and the deep temporal embedding features to the predicted load at each historical moment. The embodiments of this application can improve the accuracy of load forecasting.

Description

Technical Field

[0001] This application relates to the field of energy forecasting technology, and in particular to a two-layer load forecasting method, apparatus and equipment based on mechanism enhancement and interpretability. Background Technology

[0002] Integrated energy systems can significantly improve energy utilization efficiency and operational economy through the coordinated scheduling of multiple energy sources such as electricity, heat, and cooling. Accurate load forecasting is an important prerequisite for achieving the above-mentioned coordinated scheduling.

[0003] Currently, statistical learning methods are one of the mainstream technical solutions for load forecasting. This method mainly relies on linear modeling to build a correlation between load and historical operating data in order to achieve load forecasting.

[0004] However, in practical applications, the cooling, heating, and electrical load characteristics of integrated energy systems are complex, with various loads coupled together and deeply influenced by multiple external variables such as meteorological environment, calendar rhythms, and equipment operating status. This results in significant nonlinear variation characteristics and long-term time-series dependencies. Traditional statistical learning methods, which construct prediction frameworks based on linear modeling assumptions, can only fit simple linear relationships and are difficult to adapt to the complex load variation patterns of integrated energy systems, ultimately leading to low accuracy in load forecasting. Summary of the Invention

[0005] This application provides a two-layer load forecasting method, apparatus, and equipment based on mechanism enhancement and interpretability, which can at least improve the accuracy of load forecasting.

[0006] In a first aspect, embodiments of this application provide a two-tiered load forecasting method based on mechanism enhancement and interpretability, the method comprising: Physical functional characteristics are constructed from relevant characteristic data of the integrated energy system to obtain the corresponding physical mechanism enhancement characteristics of the integrated energy system. The physical mechanism enhancement characteristics are used to characterize the operating load of the integrated energy system in terms of physical function. The relevant characteristic data includes the historical operating load, meteorological information and calendar information of the integrated energy system. By extracting time-series features from historical operating loads, deep time-series embedding features corresponding to the integrated energy system are obtained. These deep time-series embedding features are used to characterize the time-series dependence of the operating load of the integrated energy system within historical time windows. The physical mechanism enhancement features and deep temporal embedding features are fused to obtain the fused features; Based on the fusion features, the predicted load of the integrated energy system and the corresponding interpretable information are obtained. The interpretable information is used to characterize the importance of each feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the predicted load at each historical moment.

[0007] Secondly, this application provides a two-layer load prediction device based on mechanism enhancement and interpretability, the device comprising: The construction module is used to construct physical functional characteristics from the relevant characteristic data of the integrated energy system, thereby obtaining the physical mechanism enhancement characteristics of the integrated energy system. The physical mechanism enhancement characteristics are used to characterize the operating load of the integrated energy system in terms of physical function. The relevant characteristic data includes the historical operating load of the integrated energy system, meteorological information, and calendar information. The extraction module is used to extract time-series features from historical operating loads to obtain deep time-series embedding features corresponding to the integrated energy system. The deep time-series embedding features are used to characterize the time-series dependencies of the operating load of the integrated energy system within historical time windows. The fusion module is used to fuse physical mechanism enhancement features and deep temporal embedding features to obtain fused features; The prediction module is used to make predictions based on fused features, and obtain the predicted load of the integrated energy system and the interpretable information corresponding to the predicted load. The interpretable information is used to characterize the importance of each feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the predicted load at each historical moment.

[0008] Thirdly, embodiments of this application provide an electronic device, which includes: a processor and a memory storing computer program instructions; When the processor executes computer program instructions, it implements the mechanism-enhanced and interpretable two-layer load prediction method as described in any embodiment of the first aspect.

[0009] Fourthly, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement the mechanism-enhanced and interpretable two-layer load prediction method as described in any embodiment of the first aspect.

[0010] Fifthly, embodiments of this application provide a computer program product in which instructions, when executed by a processor of an electronic device, cause the electronic device to perform the mechanism-enhanced and interpretable two-layer load prediction method as described in any of the embodiments of the first aspect above.

[0011] In the two-layer load forecasting method, apparatus, and equipment based on mechanism enhancement and interpretability provided in this application embodiment, physical functional features are constructed from relevant characteristic data such as historical operating load, meteorological information, and calendar information to generate physical mechanism enhancement features that can accurately characterize the physical operation law of the load. This can effectively adapt to the differentiated impact of various external variables such as meteorology and calendar on the load. Simultaneously, this application extracts time-series features from historical operating loads, mining deep time-series embedding features of the load within historical time windows to capture long-term time-series dependencies in load operation, solving the problem that traditional statistical methods cannot model time-series correlations. Based on this, this application fuses the enhanced features with physical mechanism characterization capabilities and deep embedding features to form fused features. Load forecasting based on these fused features can comprehensively cover the nonlinear, strongly coupled, and long-term time-series variation characteristics of the integrated energy system load, significantly reducing forecast bias and effectively improving the forecasting accuracy of the integrated energy system load. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the mechanism-enhanced and interpretable two-layer load forecasting method provided in the embodiments of this application; Figure 2 This is a schematic diagram illustrating an application scenario for predicting cooling load using the mechanism-enhanced and interpretable two-layer load prediction method provided in this application embodiment. Figure 3 This is a schematic diagram illustrating an application scenario for predicting heat load using the mechanism-enhanced and interpretable two-layer load prediction method provided in this application embodiment. Figure 4 This is a schematic diagram illustrating an application scenario for predicting electrical load using the mechanism-enhanced and interpretable two-layer load prediction method provided in this application embodiment. Figure 5 This is one of the application scenario diagrams of the two-layer load forecasting method based on mechanism enhancement and interpretability provided in the embodiments of this application; Figure 6 This is a second schematic diagram of an application scenario for the mechanism-enhanced and interpretable two-layer load forecasting method provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of a two-layer load prediction device based on mechanism enhancement and interpretability provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0015] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0016] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0017] An Integrated Energy System (IES) aims to achieve coordinated conversion and joint dispatch of multiple energy sources, such as electricity, heat, cooling, and gas, within the same park or network area. Because the multi-energy loads in this system are not independent of each other, but are deeply coupled and driven by meteorological conditions, building thermal inertia, user work and rest patterns, and the operating status of underlying equipment, they exhibit complex characteristics such as strong nonlinearity, high volatility, and intertwining of multiple time scales. Therefore, accurate multi-energy load forecasting has become a key prerequisite for carrying out system energy balance and economic operation control.

[0018] To address this need, existing multi-energy load forecasting technologies have mainly evolved into three approaches: First, statistical learning methods such as multiple regression and autoregressive moving average. These models are easy to implement, but their ability to fit high-dimensional, strongly coupled variables is severely limited. Second, traditional machine learning methods such as support vector machines and extreme gradient boosting (XGBoost). These methods possess certain nonlinear mapping capabilities, but they heavily rely on manual feature engineering and are difficult to autonomously extract global dynamic dependencies over long time series. Third, deep learning methods such as convolutional neural networks (CNN), long short-term memory networks (LSTM), and transformer models. These algorithms have significant advantages in extracting deep implicit features, but they generally suffer from inherent defects such as opaque internal decision-making mechanisms, lack of explicit physical constraints, and low engineering reliability in safety-sensitive scheduling scenarios.

[0019] Although related research has sporadically introduced underlying technologies such as Transformer's self-attention mechanism, XGBoost's nonlinear regression framework, SHapley Additive exPlanations (SHAP), and Tree-structured Parzen Estimator (TPE), showing a trend towards evolving towards a multivariate deep hybrid architecture, the above-mentioned solutions have still failed to effectively bridge the gap between prior physical mechanisms and purely data-driven models.

[0020] To address the problems existing in related technologies, embodiments of this application provide a two-layer load prediction method, apparatus, and device based on mechanism enhancement and interpretability.

[0021] The following section first introduces the two-tiered load forecasting method based on mechanism enhancement and interpretability provided in the embodiments of this application. For example... Figure 1 As shown, the method specifically includes the following steps: S101, Physical functional characteristics are constructed from the relevant characteristic data of the integrated energy system to obtain the physical mechanism enhancement characteristics corresponding to the integrated energy system; the physical mechanism enhancement characteristics are used to characterize the operating load of the integrated energy system in terms of physical function; the relevant characteristic data includes the historical operating load, meteorological information and calendar information of the integrated energy system. S102, extract time-series features from the historical operating load to obtain deep time-series embedding features corresponding to the integrated energy system; the deep time-series embedding features are used to characterize the time-series dependency of the operating load of the integrated energy system within the historical time window; S103, the physical mechanism enhancement feature and the deep temporal embedding feature are fused to obtain the fused feature; S104, Based on the fusion features, a prediction is made to obtain the predicted load of the integrated energy system and the interpretable information corresponding to the predicted load; the interpretable information is used to characterize the importance of each historical moment, as well as each feature in the physical mechanism enhancement feature and the deep temporal embedding feature, to the predicted load.

[0022] For example, in S101, an integrated energy system refers to an energy supply system that achieves coordinated conversion, joint scheduling, and complementary operation of multiple energy sources such as electricity, heat, cooling, and gas within the same park or network area. This system can include the municipal power grid, the natural gas network, and various energy conversion and utilization equipment such as gas-fired internal combustion engines and electric chillers. Its core characteristic is that different energy forms are not independent of each other, but rather deeply coupled through energy conversion equipment. Through multi-energy coordinated scheduling, the integrated energy system can significantly improve energy utilization efficiency and operational economy.

[0023] Historical operating load includes time-series measurements of cooling load, heating load, and electrical load, reflecting the actual energy demand of the integrated energy system over a past period. Meteorological information includes environmental variables such as ambient temperature, relative humidity, and solar radiation, used to characterize the impact of external weather conditions on load changes. Calendar information includes time attributes such as time and date type (weekday, weekend, holiday), used to capture the correlation between human activity rhythms and load demand. These three types of data must be aligned in the time dimension and undergo preprocessing operations such as missing value repair, outlier removal, and normalization to form standardized time-series data.

[0024] Physical mechanism enhancement features are feature vectors with clear physical meaning and business semantics, obtained by structuring and refining the original relevant feature data based on the operating mechanism of the integrated energy system.

[0025] The physical mechanism enhancement features are designed around the inherent driving forces behind the load formation of the integrated energy system, specifically encompassing periodic time encoding, calendar context, daily cold and hot values, historical statistics, lag characteristics, and future meteorological statistical features. The core function of the physical mechanism enhancement features is to compensate for the shortcomings of purely data-driven models in expressing mechanisms.

[0026] In one embodiment, the physical mechanism enhancement feature includes at least one of the following: Periodic time coding feature, which is used to characterize the periodic continuous variation pattern of the operating load of the integrated energy system at different times of day; Calendar context features, which are used to characterize the impact of different human activities on the operating load during weekdays, weekends and holidays; Daily temperature values, which are used to characterize the threshold effect of ambient temperature on the cooling and heating loads of the integrated energy system, in order to distinguish the cooling-dominant zone and the heating-dominant zone of the integrated energy system; Coupling characteristics, which are used to characterize the influence of the interaction between different influencing factors on the operating load; The statistical characteristics of the operating load within the historical time window; The hysteresis feature is used to characterize the short-term inertia and daily repetition information of the operating load; Future meteorological statistical characteristics, which are used to characterize the influence of meteorological conditions on the operating load in the future time period.

[0027] For example, the purpose of the periodic time encoding feature is to eliminate the numerical discontinuity at the daily boundary, enabling the model to perceive the periodic changes in load over time throughout the day. This feature can effectively characterize the regular fluctuation patterns of the integrated energy system's operating load at different times of the day (such as early morning, noon, evening, and night).

[0028] Calendar context features are used to characterize the impact of different human activity patterns, such as weekdays, weekends, and public holidays, on the operating load of the integrated energy system. For example, by setting binary variables such as is_weekend and is_holiday, the model is explicitly informed that under the same meteorological conditions, weekday production and office activities and rest day home life behaviors will lead to completely different cooling, heating, and electricity load demand curves. The introduction of calendar context features enables the model to adaptively adjust the forecast of load baseline levels for different date types.

[0029] Daily temperature values ​​are used to characterize the driving threshold effect of ambient temperature on the cooling and heating loads of the integrated energy system, thereby distinguishing between the cooling-dominated and heating-dominated zones of the system. This feature allows the model to intuitively identify whether the current environment has reached the critical conditions for triggering cooling or heating.

[0030] Coupling characteristics are used to characterize the combined impact of interactions between different influencing factors on the operating load of an integrated energy system. For example, multiplying temperature by a weekday identifier yields a temperature-weekday coupling characteristic, which characterizes the nonlinear enhancement effect of high temperatures on weekday conditions due to increased personnel density and equipment operation, leading to more significant cooling load demand. Multiplying temperature by relative humidity yields a temperature-humidity coupling characteristic, which reflects the additional contribution of increased latent heat load to the total cooling load under hot and humid weather.

[0031] Historical time window statistical characteristics refer to the information obtained by statistically summarizing the operating load sequence within a historical window of a preset physical duration. This information is used to characterize the recent load level, fluctuation degree, and extreme values ​​of the system. Specifically, the statistical characteristics may include the mean, standard deviation, maximum value, and minimum value.

[0032] Lag features are used to characterize the short-term inertia and diurnal repetition of operating loads. Typically, lag features may include load values ​​at the same time the previous day (t-24h, corresponding to 24 hours ago), utilizing the diurnal repetition pattern of the load. By introducing lag features with different lag periods, the model can explicitly memorize and reuse historical repetitive patterns, thereby improving its ability to predict periodic load fluctuations.

[0033] Future meteorological statistical features refer to characteristics constructed based on meteorological forecast data that reflect the overall level of meteorological conditions over a certain period of time in the future. These features characterize the impact of future meteorological conditions on the operating load of an integrated energy system. Specifically, these features may include statistical quantities such as the mean, minimum, and maximum predicted temperatures over several future steps (e.g., the next 1 hour, 6 hours, or 24 hours). Since scheduling decisions for integrated energy systems typically require forward-looking planning for a period of time, introducing future meteorological statistical features allows models to anticipate upcoming meteorological driving conditions (such as an impending cold wave or heat wave) when predicting the current load, thereby more accurately predicting the inflection points and peak levels of load changes.

[0034] In these embodiments, by constructing multi-dimensional physical mechanism enhancement features such as periodic time encoding, calendar context, daily temperature values, coupling features, historical statistics, lag and future meteorological statistics, the subsequent prediction model (such as the extreme gradient boosting tree model) can explicitly perceive the multiple driving laws of the load, thereby significantly improving the accuracy of load prediction.

[0035] In one embodiment, the physical functional feature construction of the relevant feature data of the integrated energy system to obtain the physical mechanism enhancement features corresponding to the integrated energy system includes at least one of the following: The periodic time coding feature is constructed by periodically mapping and encoding the time within the day according to the total number of sampling points per day. Construct the calendar context features according to the date type; The daily temperature values ​​are constructed based on the ambient temperature, the cooling reference temperature, and the heating reference temperature. The lag feature is constructed based on the operating load corresponding to the first moment of the first day and the operating load corresponding to the first moment of the day before the first day. The future meteorological statistical features are constructed based on the predicted temperature corresponding to the future time period.

[0036] Alternatively, in one optional implementation of this application, each physical mechanism enhancement feature can be constructed in the following manner: This application constructs physical mechanism enhancement features based on five categories of driving factors: "periodic patterns, behavioral patterns, hot and cold triggers, short-term inertia, and future boundary conditions," enabling the model to acquire a priori representation with business semantics before learning. It should be noted that the models mentioned below are all models used for subsequent load forecasting, i.e., extreme gradient boosting tree models.

[0037] First, regarding time information, this application does not directly use discrete time indices as input, but instead employs periodic mapping to eliminate numerical discontinuities at day boundaries. Let the time index within a day be... The total number of sampling points per day is Then the periodic time coding feature is:

[0038] in, The time-encoding feature corresponding to time t is used; this representation keeps adjacent period boundaries in the feature space, thus avoiding the model from misinterpreting the state near the end of the day and the beginning of the day as a sudden change.

[0039] Secondly, to characterize the structural differences in the load curve caused by the influence of human behavioral rhythms, calendar context variables such as weekdays, weekends, and holidays (i.e., calendar context features) are introduced. For example, the calendar context features corresponding to weekdays are: Calendar context features corresponding to holidays These variables do not merely label date types, but explicitly inform the model that the same meteorological conditions may correspond to completely different load responses under different activity patterns.

[0040] Furthermore, considering the significant threshold effect of cooling and heating loads on temperature-driven processes, this application introduces daily values ​​for cooling and heating (i.e., daily temperature values ​​include both cooling and heating daily values), elevating the temperature impact from a simple linear input to a more characteristic expression that aligns with HVAC triggering logic. Let the ambient temperature be... The refrigeration reference temperature is The heating reference temperature is Then we have:

[0041]

[0042] in, This is the daily value for coldness. This is the daily heat value; preferably, it is taken as... , In this way, the model can more directly distinguish whether the current environment has entered the cooling-dominated zone or the heating-dominated zone, rather than simply inferring the trigger boundary based on temperature values.

[0043] Based on this, this application further constructs coupling features, historical statistical features, lag features, and future meteorological statistical features to complete the two types of information: "multi-factor coupling" and "multi-timescale memory".

[0044] Coupling features are used to represent the coupling effect between factors such as temperature and workday attributes, and temperature and humidity, for example:

[0045]

[0046] Let be the temperature at time t, and the former can be... Characterizing the stronger driving effect of high temperature on cooling load under weekday conditions, the latter It can reflect the effect of temperature and humidity coupling leading to an increase in latent heat load, among which, Humidity t Let be the ambient humidity at time t.

[0047] Historical statistical features are used to extract summary information on load levels and fluctuation intensity over a recent period. Let the physical duration corresponding to the statistical window be... The sampling interval is Then the window length for:

[0048] At any moment Historical Window Above, the mean, standard deviation, maximum, and minimum values ​​are defined as follows:

[0049]

[0050]

[0051] in, The mean, Standard deviation This represents the actual operating load value at the i-th moment within the historical window. , These are the maximum and minimum values, respectively.

[0052] Meanwhile, to preserve short-term inertia and daily repetitive information, a hysteresis feature is constructed: ,

[0053] in, To represent the short-time inertial hysteresis characteristic at time t, the load value from the previous time step is directly introduced. This characterizes the short-term continuity of the load: the load does not change abruptly between adjacent moments, and the current load is affected by the load level of the previous moment, reflecting "short-term inertia"; To represent the daily repetitive lag characteristic at time t, the load value at the same time the previous day is introduced. This characterizes the daily periodicity of the load: the load will show similar levels at the same time every day (such as the repetition of the morning and evening peaks on weekdays), reflecting the "daily periodic repetition" pattern.

[0054] Considering that future weather forecasts are usually available for actual dispatching environments, this application also introduces future meteorological statistical features.

[0055] For example, regarding the future Step temperature forecast construction:

[0056] in, For the future The predicted temperature.

[0057] It represents the future meteorological statistical characteristics corresponding to time t, which represents the average level of predicted temperature over a future period of time; This is the predicted temperature for the h-th step in the future.

[0058] The purpose of this feature is to enable the model to anticipate the overall trend of future weather conditions, thereby providing boundary conditions to support the prediction of cold, heat, and electricity loads, and preventing the model from relying solely on historical data while ignoring future changes in operating conditions.

[0059] Therefore, the enhanced physical mechanism feature formed by this invention can be written as:

[0060] in: It is the final constructed physical mechanism enhanced feature vector. It is not a simple feature set, but a structured input built around the load formation mechanism, embedding empirical laws and physical boundaries into the model in advance.

[0061] It is a periodic time encoding feature used to characterize the periodic and continuous change pattern of time within a day.

[0062] It is a calendar context feature used to characterize the impact of different human activity patterns, such as weekdays, weekends, and holidays, on workload.

[0063] and These are the daily values ​​of cooling and heating, used to characterize the threshold effect of ambient temperature on cooling and heating loads, and to distinguish between cooling and heating-dominant areas.

[0064] It is a coupling feature used to characterize the interaction between temperature and different influencing factors such as working days and temperature and humidity.

[0065] Historical statistical characteristics, including mean, standard deviation, maximum and minimum values, are used to characterize the overall level and fluctuation characteristics of recent load.

[0066] It is a hysteresis characteristic used to characterize the short-term inertia and daily repetitive pattern of load.

[0067] These are future meteorological statistical features used to characterize the impact of meteorological conditions on load over a future period, providing future boundary conditions for the model.

[0068] This vector is not a simple set of features, but a structured input based on the load formation mechanism. Therefore, its core value lies in embedding empirical rules and physical boundaries into the prediction model in advance.

[0069] In these embodiments, explicit physical laws are efficiently embedded into the prediction model through specific construction methods such as periodic mapping encoding, date type identification, daily temperature value construction, day-ahead lag, and future meteorological statistics. This enhances the model's ability to perceive periodicity, threshold characteristics, and forward-looking driving factors, and effectively improves the accuracy of load forecasting.

[0070] In S102, deep temporal embedding features refer to the fixed-dimensional latent vector representation obtained after a model (e.g., an encoder based on a self-attention mechanism) performs multi-layer nonlinear transformations and compression encoding on the time-series sequence composed of historical operating loads. This feature is not a direct retention or simple statistical summary of the original load data, but rather a high-order abstract expression that can highly summarize the load evolution pattern within the historical time window, which is automatically learned by the model during training.

[0071] Specifically, deep temporal embedding features condense long-distance dependencies, multivariate coupling information, and dynamic evolution patterns in the original sequence into a compact vector, which serves as one of the inputs to the subsequent prediction model. This effectively compensates for the shortcomings of traditional manual feature engineering in characterizing global dependencies.

[0072] Time-series dependency refers to the correlation and constraint patterns between the operating load values ​​of an integrated energy system over a time dimension. Specifically, the load value at a given moment is not independent and random, but rather statistically correlated with its historical load values, trends, fluctuations, and the historical state of external driving factors. Time-series dependency can manifest in various forms: short-term inertia (the current load is highly correlated with the load at several recent moments), daily repetition (the similarity of the same moment across different dates), long-distance coupling (slow trends of change spanning multiple days or weeks), and abrupt correlation.

[0073] In one embodiment, the step of extracting time-series features from the historical operating load to obtain deep time-series embedded features corresponding to the integrated energy system includes: An initial representation feature is obtained by superimposing position codes on a time-series sequence composed of historical operating loads; the initial representation feature is used to characterize the correlation between the historical operating load and the time-series position at each moment in the time-series sequence. The initial representation features are processed by a multi-head attention mechanism to obtain multi-head attention; the multi-head attention is used to characterize the dependency relationship of the historical runtime load in different representation subspaces. The deep temporal embedding features are obtained by encoding the multi-head attention through a feedforward network.

[0074] Alternatively, in one optional implementation of this application, the deep temporal embedding features can be extracted in the following way: After completing the construction of physical mechanism enhancement features, this application further conducts deep temporal representation learning on historical multivariate sequences, with the goal of automatically identifying long-distance dependencies, local perturbations and multivariate coupling relationships across the entire sequence.

[0075] Integrated energy system loads exhibit significant multi-timescale characteristics. Models need to consider not only the inertial propagation of recent moments but also the impact of similar operating conditions from the previous day, previous days, and even sudden weather events on the current moment. While relying solely on traditional cyclic structures can propagate information sequentially, long-term dependency decay is prone to occur over long sequences, and there is a lack of explicit mechanisms to assign importance to different historical segments. Transformers, through their self-attention mechanism, perform global weighting on the entire historical window at each moment, making them more suitable for unified modeling of such complex dependency structures. Therefore, this application uses Transformers to extract deep temporal embedding features.

[0076] Let the input matrix be: (i.e., a time-series sequence composed of historical operating loads), where The time window length, The input dimension for a single time step.

[0077] First, the input matrix is ​​projected onto the hidden space using a linear mapping, and positional encoding is superimposed to obtain the initial representation:

[0078] in, It is the initial representation feature after linear projection and positional encoding.

[0079] This represents a linear transformation of the original time series, mapping it to the hidden dimension space required by the Transformer model, thus providing a suitable feature representation for subsequent self-attention calculations.

[0080] PE stands for Position Encoding, which is directly added to the linearly projected features to inject temporal positional information into the model. Because the Transformer itself does not have temporal order awareness, position encoding allows the model to distinguish load data at different times and understand the sequential relationship of the time series.

[0081] The core function of this step is to transform the original one-dimensional load time-series data into a high-dimensional feature vector containing time-series location information, laying the foundation for subsequent capture of long-term time-series load dependencies.

[0082] The introduction of positional encoding is necessary because the self-attention mechanism itself is insensitive to the order of inputs. For positional encoding... and dimensional index The position encoding is defined as:

[0083]

[0084] in, : These represent the values ​​of the 2i-th (even index) and 2i+1-th (odd index) dimensions at the pos-th time step in the position encoding vector, respectively.

[0085] pos: The position index of the current time in the time series.

[0086] The hidden dimension of the Transformer model is the total length of the positional encoding vector.

[0087] After positional encoding, the model can not only recognize the input content, but also distinguish the relative position of the content in the sequence.

[0088] Subsequently, in each attention head, the input representation is... Perform a linear transformation to obtain the query matrix, key matrix, and value matrix:

[0089] in, That is .

[0090] h: Head index in multi-head attention, representing the h-th attention head.

[0091] The query, key, and value weight matrix corresponding to the h-th attention head is the learnable parameter of the model.

[0092] The query vector, key vector, and value vector of the h-th attention head are used to calculate attention weights and capture temporal dependencies.

[0093] The corresponding scaled dot product attention output for:

[0094] in, The dimension of the key vector.

[0095] If adopted If there is one attention head, then multiple heads will output their own attention. for:

[0096] in, : Output of each attention head.

[0097] The concatenation operation combines the outputs of h attention heads along the feature dimension to form a longer vector.

[0098] The output projection matrix is ​​a learnable parameter of the model. Its function is to map the concatenated high-dimensional vector back to the hidden dimensions required by the model, providing a unified format of feature input for subsequent layers.

[0099] The significance of this multi-head mechanism is that different attention heads can capture different types of related patterns in parallel in different representation subspaces. For example, some heads focus more on short-term inertia, some focus more on periodic repetition, and some focus more on abrupt responses under changing weather conditions.

[0100] Following multi-head self-attention, this application continues to complete nonlinear mapping and stable training through feed-forward network (FFN), residual connections, and layer normalization.

[0101] The feedforward network FFN(z) can be written as:

[0102] in, This is the activation function.

[0103] The weight matrix and bias terms of the first-level linear transformation. It will map the feature dimension from the hidden dimension to a higher dimension (such as 4 times the hidden dimension).

[0104] The weight matrix and bias terms of the second-level linear transformation. It maps high-dimensional features back to the original hidden dimensions, providing a unified format of feature input for subsequent layers.

[0105] go through After layer encoding, the output matrix is ​​denoted as .

[0106] To compress the high-order dynamic information of the entire historical window into a fixed-length latent representation, this application performs global average pooling in the time dimension, ultimately obtaining deep temporal embedding features:

[0107] in, The final deep temporal embedding features obtained.

[0108] L: The length of the input sequence, which is the number of time steps of the historical load data.

[0109] : The feature vector at the i-th time step output by the N-th layer Transformer encoder.

[0110] Average pooling is the process of summing the feature vectors at all time points and then taking the average value.

[0111] Deep temporal embedding features essentially correspond to a compressed representation of the overall state of the historical window. They comprehensively encode the temporal dependencies and coupling relationships among multiple variables, and serve as an important implicit input for subsequent fusion regression.

[0112] In these embodiments, temporal sequence information is preserved by positional encoding, dependencies in different subspaces are captured in parallel by multi-head attention mechanism, and then nonlinear encoding by feedforward network is used to effectively extract long-distance temporal dependencies and complex coupling patterns of historical loads, which significantly improves the expressive power and prediction accuracy of deep temporal embedding features.

[0113] In S103, the fusion feature refers to the joint vector representation obtained by integrating the physical mechanism enhancement feature and the deep temporal embedding feature according to a preset combination strategy. The combination strategy includes, but is not limited to, vector concatenation, weighted summation, and weighted averaging. Taking weighted summation as an example, the fusion feature can be represented as the superposition of the two features after multiplying each by a learnable or preset weight.

[0114] The core objective of the fusion process is to enable prior knowledge of explicit physical mechanisms and dynamic abstract representations of implicit deep temporal features to work synergistically within the same representation space, forming a comprehensive input with complementary information, thereby providing more comprehensive feature support for subsequent prediction tasks. It should be noted that the specific fusion strategy adopted can be selected based on the actual application scenario and data characteristics; this application does not impose any limitations on this.

[0115] In S104, the predicted load refers to the result obtained by using a prediction model (e.g., an extreme gradient boosting tree model) to estimate the values ​​of the cooling load, heating load, and / or electrical load of the integrated energy system at a specific time or multiple future times based on the aforementioned fusion characteristics.

[0116] Load forecasting can be done in one step, which outputs the load value at a future point in time (e.g., 15 minutes in the future); or it can be done in a multi-step rolling forecast, which outputs a load sequence at multiple consecutive points in the future (e.g., every 15 minutes in the next 24 hours). The specific time granularity is adapted to the scheduling business requirements. For example, daily planning corresponds to hourly or quarter-hour resolution, intraday rolling corresponds to finer time intervals, and real-time control corresponds to minute-level or even second-level forecasting.

[0117] Interpretable information refers to a type of auxiliary output generated along with the output predicted load, which reveals the basis for the predictive model's decision-making. The interpretable information includes at least two dimensions: time-dimensional interpretable information and feature-dimensional interpretable information.

[0118] The time dimension interpretation information is used to quantify the contribution of each historical moment to the current prediction load. For example, the attention weight matrix generated during the temporal feature extraction process gives the importance weight of each historical moment to answer "which historical moments the model focuses on". The feature dimension interpretation information is used to quantify the marginal contribution of each specific feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the prediction load. For example, the contribution value of each feature is calculated by the Shapley additive interpretation method to answer "which input features promote or inhibit the prediction results".

[0119] The introduction of interpretable information makes the internal workings of the prediction model explicit, enhancing the dispatchers' confidence in the model's output and thus meeting the interpretability requirements of safety-sensitive scenarios such as integrated energy systems.

[0120] In one embodiment, the step of predicting the load of the integrated energy system based on the fusion features includes: The fused features are input into an extreme gradient boosting tree model, and the prediction load is obtained by making a prediction using the extreme gradient boosting tree model. In the training process of the extreme gradient boosting tree model, hyperparameter optimization is performed using a Bayesian optimization method based on a tree structure Pazen estimator.

[0121] Optionally, in one feasible implementation of this application, the predicted load can be obtained by the following method: After obtaining the physical mechanism enhancement features and deep temporal embedding features, this application enters the prediction and decision-making stage. This stage is designed not to have the Transformer directly output the final payload, but rather to use XGBoost (Extreme Gradient Boosting Tree) to perform nonlinear regression after feature fusion. The reason for this design is that while the Transformer excels at extracting high-dimensional latent states from temporal data, its direct regression output often still retains strong black-box properties; whereas XGBoost is highly adaptable to heterogeneous feature inputs, local conditional splits, and nonlinear interactions, making it particularly suitable for handling input scenarios where "explicit physical features and implicit deep features coexist."

[0122] Therefore, this application integrates the two types of information in the regression space to simultaneously utilize the deep representation capabilities and the robust decision-making capabilities of the tree model.

[0123] Let the deep temporal embedding features output in the first stage be... The physical mechanism enhances the characteristics. The fused feature vector is then defined as:

[0124] As a feature fusion vector, it retains both the potential generalization of historical operating trajectories and the explicit description of current and future boundary conditions.

[0125] For the regressor, this invention uses XGBoost.

[0126] Suppose that An ensemble model consisting of 10 regression trees is denoted as . The predicted value is:

[0127] in, This represents the set of regression tree functions.

[0128] : The model's prediction for time t+1.

[0129] : The predicted output of the k-th regression tree.

[0130] In multi-step prediction scenarios, this application employs an iterative autoregressive rolling mechanism to extend the single-step model to multiple future time steps. For the... The prediction of a future moment is in the form of:

[0131] in, : The model's prediction for the k-th future time.

[0132] The completed prediction model.

[0133] The input sequence of the model, i.e., the historical data window used to predict future moments. Where: L: The length of the input sequence; t+kL: The start time of the input window; t+k-1: The end time of the input window; The fixed set of parameters after the model has been trained.

[0134] The external impact data required for the dynamic forecast window are directly derived from known information such as weather forecasts; subsequent load values ​​are gradually calculated and updated based on the results of the previous forecast. The significance of this mechanism is that it enables the model to continuously adjust its forecast trajectory by incorporating new boundary conditions during the rolling process, rather than mechanically continuing historical load curves. Therefore, it is more suitable for supporting day-ahead and intraday coordinated scheduling scenarios.

[0135] The model training objective is defined as:

[0136] in, For the loss function, mean squared error is preferred; The objective function is the total loss that needs to be optimized during model training.

[0137] n: Total number of samples; : The true value of the i-th sample; The predicted value of the i-th sample; : The k-th regression tree; Complexity penalty term:

[0138] in, For the first The number of leaf nodes of a tree The leaf node weight vector, and This is the regularization coefficient. This regularization term is used to limit the complexity of the tree structure and prevent the model from overfitting to local noise.

[0139] To achieve efficient training, XGBoost employs a second-order approximation of the objective function. Let the... The new tree is added in round Then the objective function It can be written as:

[0140] in: The output prediction increment for the t-th tree on the i-th sample;

[0141] in, The first-order partial derivative of the loss function for the i-th sample; The second-order partial derivative of the loss function for the i-th sample; Single-sample loss function, which measures the error between the actual load and the predicted load; : The prediction load of the model for the i-th sample in the (t-1)th iteration.

[0142] With this approach, the model can efficiently complete the search for the optimal split point and the calculation of leaf node weights in each iteration.

[0143] Considering that XGBoost performance is highly sensitive to parameter selection, this application further introduces a tree-based Pazen estimator (Tree). Hyperparameter optimization is performed using Bayesian optimization of the structured Parzen Estimator (TPE). Let the hyperparameter vector be... The objective function of the validation set is Then optimize the objective for:

[0144] in, Given hyperparameters The loss of the model on the validation set.

[0145] In its implementation, TPE divides the parameter space into good-value regions and non-good-value regions, and models them separately:

[0146]

[0147] Where y represents the model's hyperparameters The following performance indicators To verify the error threshold, p is the conditional probability.

[0148] Subsequently, the algorithm prioritizes... A larger combination of parameters improves search efficiency. Therefore, this application not only presents the feature fusion and regression decision structure, but also an optimization mechanism adapted to this structure, enabling the model to maintain good generalization performance under different load types and operating conditions.

[0149] To ensure this application has a complete engineering implementation path, it is also necessary to provide a unified explanation of model training, performance evaluation, and system linkage methods. During the training phase, the mean squared error loss function is preferred for single-step regression tasks. :

[0150] This loss function is more sensitive to samples with large deviations, thus it helps to suppress prediction errors during peak periods and when operating conditions change, where N is the total number of samples. This represents the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0151] After training, this application uses the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) to jointly evaluate the model performance. Let the number of test samples be... The true mean is ,but:

[0152]

[0153]

[0154]

[0155] in, Used to emphasize the penalty for large errors Reflects the average deviation level, Suitable for comparing the relative errors between different load types. This describes the model's ability to explain real fluctuations. By using multiple indicators for joint evaluation, we can more comprehensively assess the model's performance in different application scenarios, rather than relying on a single indicator to draw one-sided conclusions.

[0156] In one specific embodiment, such as Figure 2 , Figure 3 , Figure 4 As shown, this application has been verified through experiments and engineering scenario data, using a comprehensive energy system in an industrial park in a certain city as the experimental verification object. This system provides integrated energy supply of cooling, heating, and electricity to an office building complex of approximately 30,000 square meters. Its energy side includes municipal electricity, natural gas network, and key equipment such as gas internal combustion engines, electric chillers, absorption chillers, and gas boilers, exhibiting typical multi-energy coupling characteristics.

[0157] The load data used in the experiment originated from real-time measurements of the park's data acquisition and monitoring control system, with a sampling resolution set at 15 minutes, spanning from January 1, 2023 to December 31, 2024. Simultaneously collected meteorological data included environmental variables such as temperature, relative humidity, and solar radiation. The overall experimental dataset was strictly divided into training, validation, and test sets in a 7:1:2 ratio according to the time series.

[0158] Experiments conducted based on the aforementioned fundamental data have demonstrated that the method described in this application possesses extremely high accuracy in predicting three types of loads: cooling, heating, and electricity. The prediction results for the three types of loads are as follows: Figure 2 , Figure 3 , Figure 4 As shown in the figure. Specific quantitative evaluation indicators show that: the predicted RMSE for cooling load is 52.53kW, MAE is 31.03kW, MAPE is 3.47%, and R² is 0.9976; the predicted RMSE for heating load is 83.41kW, MAE is 63.2kW, MAPE is 7.378%, and R² is 0.9787; the predicted RMSE for electrical load is 14.6kW, MAE is 9.38kW, MAPE is 3.58%, and R² is 0.9882.

[0159] In one embodiment, the interpretable information includes: First interpretable information, which characterizes the importance of each historical moment to the predicted load; The deep temporal embedding features are extracted by a Transformer encoder. The first interpretable information is determined based on the attention weight matrix of the Transformer encoder. Each element in the attention weight matrix is ​​used to characterize the correlation strength between the prediction time corresponding to the prediction load and each historical time.

[0160] In one embodiment, the interpretable information includes: The second interpretable information is the marginal contribution of each feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the predicted load; The prediction process of the predicted load is analyzed using the Shapley additive interpretation method to obtain the marginal contribution of each feature. When the marginal contribution is positive, the corresponding feature has a positive driving effect on the predicted load; when the marginal contribution is negative, the corresponding feature has an inhibitory effect on the predicted load.

[0161] Optionally, in one feasible implementation of this application, the first interpretable information and the second interpretable information can be determined in, but are not limited to, the following ways: In addition to providing predicted values, this application also constructs a two-layer interpretable output mechanism to improve the reliability of the model in integrated energy system scheduling applications.

[0162] The first layer of explanation (i.e., the first interpretable information) comes from the attention weights within the Transformer, which answer the question of "which historical moments the model primarily references in the time dimension".

[0163] Let the attention matrix be Then its elements are defined as:

[0164] in, The element in the i-th row and j-th column of the attention weight matrix represents the attention weight of the i-th position to the j-th position.

[0165] Q: Query matrix.

[0166] K: Key matrix.

[0167] The vector dimensions of Q and K are described in the foregoing embodiments.

[0168] By visualizing the attention distribution corresponding to the prediction time, an importance heatmap of historical time steps can be obtained, thereby identifying whether the model relies more on short-term inertia, daily cycle patterns, or certain anomalous fluctuation segments in the current prediction. This explanatory mechanism enables the model to not only output results but also explain "why these historical intervals are being focused on".

[0169] The second layer of explanation (i.e., the second interpretable information) comes from the SHAP analysis in the XGBoost output stage, which answers the question "which input features drove the predicted value up or down?" For any input sample The model output f(x) can be written as:

[0170] in, As the baseline output, For the first The marginal contribution of each feature; M is the number of independent feature components, i.e., the total number of basis functions. Based on cooperative game theory, the features... The SHAP value is defined as:

[0171] in, The SHAP value; For the complete feature set, For features not included Any subset of features; Objective function (the model's performance metrics under these hyperparameters, such as validation set loss). If This indicates that the feature has a positive driving effect on the current sample; if This indicates that the feature has a suppressive effect on the predicted value.

[0172] Because the input of this application includes both physical mechanism enhancement features and deep temporal embedding features, SHAP can not only explain the contributions of explicit physical factors such as temperature, humidity, daily coldness, and daily heatness, but also quantify the influence of latent deep representations (i.e., deep temporal embedding features) in the final decision. This means that the model can provide explanations in both the time and feature dimensions, thus forming a complete two-layer interpretable framework.

[0173] In these embodiments, the importance of historical moments in the time dimension is explained by the attention weight matrix, and the marginal contribution of each feature is quantified by the Shapley additive interpretation method, achieving a two-layer interpretable output in both the time and feature dimensions, which significantly enhances the decision transparency and engineering credibility of the prediction model.

[0174] In one embodiment, the method further includes: Based on the predicted load, and with the equipment output constraints, unit start-up and shutdown constraints, system energy balance constraints, and operating costs of the integrated energy system as optimization objectives, the unit start-up and shutdown plan and equipment output setpoints of the integrated energy system are generated. The equipment output constraint is used to limit the upper and lower limits of the normal output of various energy equipment in the integrated energy system. The unit start-stop constraints are used to regulate the start-stop logic and switching restrictions of the energy units in the integrated energy system. The system energy balance constraint is used to maintain a balance between the total energy supply of the integrated energy system and the total energy consumption on the user side.

[0175] For example, equipment output constraints refer to conditions that limit the allowable output range of various energy devices in an integrated energy system under normal operating conditions. Specifically, each type of equipment has a minimum and maximum output value determined by its technical characteristics, and equipment output constraints require that the actual output of each device must fall within this closed range. This constraint is used to ensure that the equipment operates within a safe and efficient operating range, avoiding equipment damage or energy efficiency degradation due to overload or underload.

[0176] Unit start-up and shutdown constraints refer to a set of rules used to regulate the start-up and shutdown logic and state transition restrictions of various energy units in an integrated energy system. These constraints include, but are not limited to: minimum operating time (the shortest duration a unit must run continuously after startup), minimum downtime (the shortest duration a unit must remain in an out-of-operation state after shutdown), start-up and shutdown frequency limits (the maximum number of start-ups and shutdowns allowed within a given time period), and logical coordination of start-up and shutdown sequences. The introduction of unit start-up and shutdown constraints aims to avoid problems such as accelerated equipment wear, increased energy loss, and decreased system stability caused by frequent start-ups and shutdowns.

[0177] System energy balance constraints refer to the requirement that, at any given time, the total energy supply of all types of energy in an integrated energy system must remain equal to the total energy consumption by users. Specifically, for each form of energy—electricity, heat, and cooling—the sum of energy generated by all equipment within the system, minus the energy consumed by the system itself, must equal the energy demand of users at that moment. This constraint is a core physical condition for ensuring stable system operation; violating it will lead to operational risks such as frequency fluctuations or insufficient heating / cooling.

[0178] A unit start-up and shutdown plan is a scheduling scheme that pre-arranges the operating status of each energy unit in an integrated energy system over time, based on load forecasting results and system optimization objectives. This plan clearly indicates, in time-series form, when each unit will start and when it will shut down. The unit start-up and shutdown plan is one of the core outputs of day-ahead or intraday scheduling decisions in an integrated energy system, and its rationality directly affects the system's operational economy, reliability, and equipment lifespan.

[0179] Equipment output setpoints refer to the specific power values ​​that each operating energy device in an integrated energy system should output at a specific dispatch time, determined based on the unit start-up and shutdown plan and system optimization objectives. Equipment output setpoints and unit start-up and shutdown plans together constitute a complete dispatch instruction. The start-up and shutdown plan determines whether the equipment is operating, while the output setpoint determines the actual load level of the operating equipment. Together, they work at the system control execution layer to achieve refined operation and control of the integrated energy system.

[0180] Optionally, in one specific implementation of this application, such as Figure 5 As stated above, at the engineering deployment level, this application embeds the prediction model into the integrated energy system monitoring and optimization scheduling platform, placing it at the data hub position between the bottom-level data acquisition and the upper-level decision control.

[0181] Specifically, the data acquisition module continuously provides historical load, real-time operating status, and future weather information; the forecasting module outputs load forecast results at three time scales: day-ahead, intraday, and real-time, based on the above inputs and outputs; the optimization scheduling engine generates unit start-up and shutdown plans, output setpoints, and demand response strategies based on the forecast results, combined with equipment output range, start-up and shutdown constraints, energy balance constraints, and operating cost targets; and the control execution layer implements corresponding controls based on the scheduling results.

[0182] Furthermore, at the practical engineering implementation level, the method described in this application can serve as a core bridge between the underlying multi-energy data acquisition module and the upper-level optimization control module, seamlessly integrating its prediction results into the integrated energy system monitoring and optimization scheduling platform. This mechanism supports the output of multi-energy load prediction results at three time scales: day-ahead, intraday, and real-time. The day-ahead prediction data can provide a decision-making basis for the formulation of unit combination and multi-energy coordinated scheduling schemes; the high-frequency real-time prediction data effectively supports dynamic operation and adjustment operations such as intraday rolling optimization, supply and demand deviation correction, and demand response management, thus demonstrating significant engineering implementation and application value.

[0183] Therefore, the prediction model in this application is no longer a simple offline analysis tool, but a key upstream module in the closed-loop operation of the integrated energy system.

[0184] Furthermore, to address seasonal changes, equipment status drift, and long-term operating condition evolution, this application also allows for model updates to be triggered based on prediction error drift, thereby maintaining prediction accuracy and scheduling reliability during long-term operation.

[0185] In these embodiments, based on predicted load, multiple constraints such as equipment output, unit start-up and shutdown, and energy balance are integrated, and scheduling schemes are generated with operating costs as the optimization objective. This achieves closed-loop linkage between prediction and scheduling, effectively improving the operational economy and reliability of the integrated energy system.

[0186] It should be noted that the various optional implementation methods described in the embodiments of this application can be combined with each other or implemented individually without conflict, and the embodiments of this application do not limit this.

[0187] In a complete embodiment, this application addresses the characteristic that the cooling load, heating load, and electrical load in an integrated energy system are simultaneously affected by historical operating conditions, meteorological boundary conditions, time cycle patterns, and user behavior patterns. It constructs an overall technical approach of "physical mechanism enhanced feature construction + deep temporal embedding feature extraction + XGBoost fusion regression decision + two-layer interpretable output + rolling prediction and scheduling linkage".

[0188] like Figure 6 As shown. The core idea of ​​this technical approach is not to directly feed multi-source data into a single model for end-to-end fitting, but rather to first extract explicit features with clear business and physical significance from the raw data, enabling the model to perceive cold and heat driving thresholds, time periodicity, operational inertia, and future meteorological boundaries (i.e., physical mechanism enhancement features). Second, it uses Transformer to automatically extract high-dimensional potential dynamic representations from historical multivariate sequences to capture long-term dependencies, local mutations, and cross-variable coupling relationships (i.e., deep temporal embedding features). Then, it embeds the physical mechanism enhancement features and deep temporal embedding features in a unified space and fuses them, and completes the final nonlinear regression prediction through XGBoost. Furthermore, it simultaneously provides time dimension interpretation results based on attention weights and feature dimension interpretation results based on SHAP at the model output stage, thereby enhancing the transparency of the prediction process. Finally, it extends single-step prediction to multi-step rolling prediction and sends the prediction results to the integrated energy system optimization and scheduling platform to support day-ahead scheduling, intraday rolling optimization, and real-time operation control.

[0189] Based on this technical approach, this application actually forms a complete technology chain from data representation, mechanism embedding, dynamic learning, predictive decision-making to scheduling applications, rather than a simple patchwork of several isolated algorithm modules. Therefore, this application achieves a complete chain from data, features, models, interpretation to applications, improving both the accuracy and robustness of multi-energy load forecasting, and enhancing model transparency and engineering feasibility.

[0190] Specifically, this application adopts a two-stage overall modeling mechanism to explicitly decouple deep temporal representation learning from the final regression decision.

[0191] Set time The target load is The exogenous covariate vector is The unified observation vector is defined as follows: Among them, the exogenous covariate vector refers to external influencing factors that affect the load but are not determined by the load's own laws.

[0192] Given a length of The historical observation window, the input sequence is denoted as The predicted step size is At that time, the future load sequence is denoted as .

[0193] In the single-step prediction scenario, the core prediction mapping of this application is represented as follows:

[0194] in, Represents deep temporal embedding features. Indicates enhanced physical mechanism characteristics. This represents a nonlinear regression mapping based on XGBoost. Represents the set of model parameters; This is the predicted value at time t+1.

[0195] This formula shows that the final predicted value is not directly generated by a single path, but is jointly determined by "deep temporal embedding features" and "physical mechanism enhancement features". Thus, this application achieves collaborative modeling of "explicit mechanism information" and "implicit deep information" at the structural level.

[0196] Compared with related technologies, this application proposes a two-stage decoupled prediction architecture that balances high prediction accuracy and high engineering transparency. In terms of model structure, this application abandons the traditional black-box approach of directly outputting prediction results from a single deep network, and explicitly decouples deep temporal representation learning from nonlinear regression decision-making.

[0197] Specifically, the Transformer encoder extracts high-order complex dynamic dependencies in the sequence, and the XGBoost model performs nonlinear mapping on the fused features, thereby effectively reducing the difficulty of deep networks independently undertaking the full-link fitting task, and taking into account both the global temporal modeling capability and the robust decision-making performance of tree models when dealing with heterogeneous features.

[0198] In terms of feature representation, compared with schemes that rely solely on manual features or deep representations, this application achieves a complementary fusion of explicit physical mechanisms and implicit dynamic information. While retaining explicit physical features such as time attributes, heating and cooling sources, historical statistics, and future weather conditions, the model introduces deep temporal embedding vectors, thereby enabling more accurate discovery of the complex coupling evolution patterns between cold, heat, and electrical loads.

[0199] Regarding model interpretability, this application prioritizes transparency design as a core structural element, constructing a two-layer interpretability mechanism encompassing both time and features. By leveraging attention weights to analyze key dependencies in the time dimension and combining SHAP values ​​to quantify the global and local contributions of each feature dimension, it directly reveals the historical moments and dominant physical driving factors that the model focuses on during decision-making, overcoming the limitation of traditional hybrid models that can only output predicted values.

[0200] At the engineering deployment level, this application extends from offline verification to real-world business processes, constructing a complete closed loop encompassing confidence interval output, multi-timescale prediction, and standardized scheduling interfaces. This solution serves as a key hub for underlying state monitoring and upper-level optimization control, supporting day-ahead planning, intraday rolling optimization, and real-time forecast correction. It can seamlessly integrate with the comprehensive energy system optimization and scheduling platform, thereby providing operators with highly reliable decision support.

[0201] Figure 7 A schematic diagram of a two-layer load prediction device based on mechanism enhancement and interpretability provided in another embodiment of this application is shown. For ease of explanation, only the parts related to the embodiments of this application are shown.

[0202] Reference Figure 7 A mechanism-enhanced and interpretable two-tier load forecasting device may include: The construction module 701 is used to construct physical functional characteristics from the relevant characteristic data of the integrated energy system to obtain the physical mechanism enhancement characteristics of the integrated energy system. The physical mechanism enhancement characteristics are used to characterize the operating load of the integrated energy system in terms of physical function. The relevant characteristic data includes the historical operating load of the integrated energy system, meteorological information and calendar information. The extraction module 702 is used to extract time-series features from historical operating loads to obtain deep time-series embedding features corresponding to the integrated energy system; the deep time-series embedding features are used to characterize the time-series dependency of the operating load of the integrated energy system within the historical time window; The fusion module 703 is used to fuse physical mechanism enhancement features and deep temporal embedding features to obtain fused features; The prediction module 704 is used to make predictions based on fused features to obtain the predicted load of the integrated energy system and the interpretable information corresponding to the predicted load. The interpretable information is used to characterize the importance of each feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the predicted load at each historical moment.

[0203] Figure 8 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0204] The device may include a processor 801 and a memory 802 storing program instructions.

[0205] When processor 801 executes the program, it implements the steps in any of the above method embodiments.

[0206] For example, the program can be divided into one or more modules / units, one or more of which are stored in memory 802 and executed by processor 801 to complete this application. The one or more modules / units can be a series of program instruction segments capable of performing a specific function, which describe the execution process of the program in the device.

[0207] Specifically, the processor 801 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0208] Memory 802 may include mass storage for data or instructions. For example, and not limitingly, memory 802 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 802 may include removable or non-removable (or fixed) media. Where appropriate, memory 802 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 802 is non-volatile solid-state memory.

[0209] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the methods according to one aspect of this disclosure.

[0210] The processor 801 implements any of the methods described in the above embodiments by reading and executing program instructions stored in the memory 802.

[0211] In one example, the electronic device may also include a communication interface 803 and a bus 810. The processor 801, memory 802, and communication interface 803 are connected via the bus 810 and communicate with each other.

[0212] The communication interface 803 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0213] Bus 810 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCIe) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 810 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0214] Furthermore, in conjunction with the methods in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores program instructions; when these program instructions are executed by a processor, they implement any of the methods in the above embodiments.

[0215] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0216] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0217] This application provides a computer program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above method embodiments and achieve the same technical effects. To avoid repetition, it will not be described again here.

[0218] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0219] The functional modules shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0220] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0221] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to create a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0222] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A two-tiered load forecasting method based on mechanism enhancement and interpretability, characterized in that, include: Physical functional features are constructed from relevant characteristic data of the integrated energy system to obtain the physical mechanism enhancement features corresponding to the integrated energy system. The physical mechanism enhancement features are used to characterize the operating load of the integrated energy system in terms of physical function; the relevant feature data include the historical operating load of the integrated energy system, meteorological information, and calendar information. The historical operating load is subjected to time-series feature extraction to obtain the deep time-series embedding features corresponding to the integrated energy system; The deep temporal embedding feature is used to characterize the temporal dependency of the operating load of the integrated energy system within a historical time window; The physical mechanism enhancement features and the deep temporal embedding features are fused to obtain the fused features; Based on the fusion features, a prediction is made to obtain the predicted load of the integrated energy system and the interpretable information corresponding to the predicted load; the interpretable information is used to characterize the importance of each historical moment, as well as each feature in the physical mechanism enhancement feature and the deep temporal embedding feature, to the predicted load.

2. The method according to claim 1, characterized in that, The physical mechanism enhancement feature includes at least one of the following: Periodic time coding feature, which is used to characterize the periodic continuous variation pattern of the operating load of the integrated energy system at different times of day; Calendar context features, which are used to characterize the impact of different human activities on the operating load during weekdays, weekends and holidays; Daily temperature values, which are used to characterize the threshold effect of ambient temperature on the cooling and heating loads of the integrated energy system, in order to distinguish the cooling-dominant zone and the heating-dominant zone of the integrated energy system; Coupling characteristics, which are used to characterize the influence of the interaction between different influencing factors on the operating load; The statistical characteristics of the operating load within the historical time window; The hysteresis feature is used to characterize the short-term inertia and daily repetition information of the operating load; Future meteorological statistical characteristics, which are used to characterize the influence of meteorological conditions on the operating load in the future time period.

3. The method according to claim 2, characterized in that, The physical functional feature construction of the relevant characteristic data of the integrated energy system to obtain the physical mechanism enhancement features corresponding to the integrated energy system includes at least one of the following: The periodic time coding feature is constructed by periodically mapping and encoding the time within the day according to the total number of sampling points per day. Construct the calendar context features according to the date type; The daily temperature values ​​are constructed based on the ambient temperature, the cooling reference temperature, and the heating reference temperature. The lag feature is constructed based on the operating load corresponding to the first moment of the first day and the operating load corresponding to the first moment of the day before the first day. The future meteorological statistical features are constructed based on the predicted temperature corresponding to the future time period.

4. The method according to claim 1, characterized in that, The step of extracting time-series features from the historical operating load to obtain the deep time-series embedded features corresponding to the integrated energy system includes: An initial representation feature is obtained by superimposing position codes on a time-series sequence composed of historical operating loads; the initial representation feature is used to characterize the correlation between the historical operating load and the time-series position at each moment in the time-series sequence. The initial representation features are processed by a multi-head attention mechanism to obtain multi-head attention; the multi-head attention is used to characterize the dependency relationship of the historical runtime load in different representation subspaces. The deep temporal embedding features are obtained by encoding the multi-head attention through a feedforward network.

5. The method according to claim 1, characterized in that, The prediction based on the fusion characteristics to obtain the predicted load of the integrated energy system includes: The fused features are input into an extreme gradient boosting tree model, and the prediction load is obtained by making a prediction using the extreme gradient boosting tree model. In the training process of the extreme gradient boosting tree model, hyperparameter optimization is performed using a Bayesian optimization method based on a tree structure Pazen estimator.

6. The method according to claim 1, characterized in that, The interpretable information includes: First interpretable information, which characterizes the importance of each historical moment to the predicted load; The deep temporal embedding features are extracted by a Transformer encoder. The first interpretable information is determined based on the attention weight matrix of the Transformer encoder. Each element in the attention weight matrix is ​​used to characterize the correlation strength between the prediction time corresponding to the prediction load and each historical time.

7. The method according to claim 1, characterized in that, The interpretable information includes: The second interpretable information is the marginal contribution of each feature in the physical mechanism enhancement feature and the deep temporal embedding feature to the predicted load; The prediction process of the predicted load is analyzed using the Shapley additive interpretation method to obtain the marginal contribution of each feature. When the marginal contribution is positive, the corresponding feature has a positive driving effect on the predicted load; when the marginal contribution is negative, the corresponding feature has an inhibitory effect on the predicted load.

8. The method according to claim 1, characterized in that, The method further includes: Based on the predicted load, and with the equipment output constraints, unit start-up and shutdown constraints, system energy balance constraints, and operating costs of the integrated energy system as optimization objectives, the unit start-up and shutdown plan and equipment output setpoints of the integrated energy system are generated. The equipment output constraint is used to limit the upper and lower limits of the normal output of various energy equipment in the integrated energy system. The unit start-stop constraints are used to regulate the start-stop logic and switching restrictions of the energy units in the integrated energy system. The system energy balance constraint is used to maintain a balance between the total energy supply of the integrated energy system and the total energy consumption on the user side.

9. A two-layer load prediction device based on mechanism enhancement and interpretability, characterized in that, The device includes: A construction module is used to construct physical functional characteristics from relevant characteristic data of the integrated energy system to obtain the physical mechanism enhancement characteristics corresponding to the integrated energy system; the physical mechanism enhancement characteristics are used to characterize the operating load of the integrated energy system in terms of physical function; the relevant characteristic data includes the historical operating load, meteorological information and calendar information of the integrated energy system; An extraction module is used to extract time-series features from the historical operating load to obtain deep time-series embedding features corresponding to the integrated energy system; the deep time-series embedding features are used to characterize the time-series dependency of the operating load of the integrated energy system within a historical time window; The fusion module is used to fuse the physical mechanism enhancement features and the deep temporal embedding features to obtain fused features; The prediction module is used to make predictions based on the fusion features to obtain the predicted load of the integrated energy system and the interpretable information corresponding to the predicted load; the interpretable information is used to characterize the importance of each historical moment, as well as each feature in the physical mechanism enhancement feature and the deep temporal embedding feature, to the predicted load.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the two-layer load forecasting method based on mechanism enhancement and interpretability as described in any one of claims 1 to 8.

Images