Power saving method based on enterprise electricity consumption power load prediction

Abstract

The invention relates to the technical field of power load prediction, in particular to a power saving method based on enterprise power consumption power load prediction, which comprises the following steps: acquiring historical power load time sequence data and historical power market data of an enterprise; performing feature extraction on the acquired data to obtain a first feature set and a second feature set; combining the first feature set with the second feature set to generate a fused feature set; an enterprise power load prediction model is constructed, and training of the enterprise power load prediction model is completed through fusion of the feature set; and solving the fusion feature set obtained in real time based on the trained enterprise power load prediction model, and solving an optimal power saving strategy according to an enterprise power load prediction value in combination with a defined objective function. According to the method, the load prediction precision is improved through the AI model fusing the physical information, the operation cost and the prediction accuracy are balanced by applying the multi-objective optimization algorithm, and the enterprise energy management level is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a power saving method based on enterprise power load forecasting. Background Art

[0002] With the acceleration of industrialization and the transformation of energy structures, enterprise-level energy management faces increasingly severe challenges and opportunities. Electricity, as the core driving force of modern industrial production, accounts for a significant proportion of a company's total operating expenses. Furthermore, the increasing popularity of mechanisms such as time-of-use pricing and demand response has placed higher demands on enterprises for refined energy management. Currently, traditional methods used for enterprise load forecasting and management have inherent limitations. First, many traditional forecasting models, such as time series analysis or simple regression, fail to fully capture the complex nonlinear dynamics within industrial loads and the strong coupling between equipment operation and external environmental factors. This results in a significant decrease in forecast accuracy, especially when loads fluctuate significantly or operating conditions change. Second, traditional energy-saving strategies are often based on empirical rules or simple threshold controls. They lack accurate foresight of future load trends and electricity price fluctuations, often employing a reactive response approach rather than proactive planning. This makes it difficult to systematically explore potential energy savings, limiting overall energy management effectiveness. Summary of the Invention

[0003] The purpose of the present invention is to provide an energy-saving method based on enterprise electricity load forecasting. It improves the load forecasting accuracy by integrating AI models with physical information, and uses multi-objective optimization algorithms to balance multiple objectives such as operating costs and forecast accuracy while meeting various constraints, thereby formulating a more accurate, economical and robust optimal energy-saving strategy, thereby significantly improving the enterprise's energy management level.

[0004] The present invention is achieved through the following technical solutions:

[0005] The power saving method based on enterprise power load forecasting includes the following steps:

[0006] Obtain the company's historical power load time series data and historical power market data;

[0007] Perform feature extraction on the enterprise's historical power load time series data and the historical power market data respectively to obtain a first feature set corresponding to the enterprise's historical power load time series data and a second feature set corresponding to the historical power market data;

[0008] Combining the first feature set with the second feature set to generate a fused feature set;

[0009] Build an enterprise power load forecasting model based on the PINN-SAHM neural network, input the fused feature set into the enterprise power load forecasting model for calculation until the training of the enterprise power load forecasting model is completed;

[0010] The fusion feature set obtained in real time is solved based on the trained enterprise power load forecasting model. The optimal power-saving strategy is solved according to the enterprise power load forecast value and the defined objective function.

[0011] Optionally, the acquisition of the enterprise's historical power load time series data and historical power market data further includes a preprocessing process, which is specifically as follows:

[0012] Fill missing values and smooth outliers in the company's historical power load time series data;

[0013] Standardize historical electricity market data;

[0014] And align the time stamps of the enterprise's historical power load time series data with the corresponding historical power market data;

[0015] Complete data preprocessing of the company's historical power load time series data and historical power market data.

[0016] Optionally, the feature extraction is performed on the enterprise's historical power load time series data and the historical power market data respectively, wherein the feature extraction is performed on the enterprise's historical power load time series data, and the specific steps are:

[0017] Applying a predetermined wavelet mother function to the enterprise's historical power load time series data to perform multi-layer wavelet packet decomposition to obtain multiple wavelet packet node signals covering different frequency bands;

[0018] Calculating a characteristic quantity of at least a portion of the wavelet packet node signal, wherein the characteristic quantity is selected from at least one of energy, entropy, variance, skewness, or kurtosis;

[0019] The calculated feature quantities are combined into the first feature set.

[0020] Optionally, feature extraction is performed on historical electricity market data. The specific steps are as follows:

[0021] Identify key market events of predefined types and their occurrence times in historical electricity market data;

[0022] For the identified key market events, extract the historical power load subsequence within the corresponding time window;

[0023] The historical power load subsequences are processed based on the symbolic aggregation approximation algorithm to generate a symbol string representing the load pattern;

[0024] Mining the frequent association patterns between key market events and generated load pattern symbol strings;

[0025] The second feature set is constructed based on the mined frequent association patterns.

[0026] Optionally, the first feature set and the second feature set are combined as follows:

[0027] representing the first feature set as a first numerical feature vector;

[0028] representing the second feature set as a second numerical feature vector;

[0029] Sequentially appending the second numerical feature vector to the first numerical feature vector to form a single fused feature vector whose dimension is the sum of the dimensions of the first and second numerical feature vectors;

[0030] The single fused feature vector is determined as the fused feature set, and the fused feature set is input into the enterprise power load forecasting model.

[0031] Optionally, the enterprise power load forecasting model is constructed based on the PINN-SAHM neural network, which is specifically:

[0032] An input layer of the enterprise power load forecasting model is defined and configured to receive the fused feature set as input;

[0033] An RNN layer is used to process the sequence output by the input layer to capture the temporal dependency in the fused feature set and output an RNN output sequence containing the hidden state of each time step;

[0034] a self-attention mechanism layer, configured to receive the RNN output sequence as input, calculate attention weights between hidden states at different time steps in the RNN output sequence, and perform a weighted summation of the RNN output sequence based on the attention weights to generate a context vector;

[0035] The MLP layer receives the context vector as input, performs nonlinear transformation through at least one fully connected layer, and outputs the enterprise power load forecast value.

[0036] Optionally, the training process of the enterprise power load forecasting model is:

[0037] Construct a training data set, where each training sample contains a fusion feature set corresponding to a historical time period and the corresponding actual enterprise power load value;

[0038] Define the total loss function of the enterprise power load forecasting model. The total loss function consists of a data-driven loss term and a physical information constraint loss term, and is weighted and summed by a preset weight λ. The data-driven loss term is specifically the mean square error between the predicted enterprise power load value output by the enterprise power load forecasting model and the corresponding actual enterprise power load value in the training sample; the physical information constraint loss term is specifically the mean square value of the residual of the load change rate constraint equation obtained by substituting the predicted enterprise power load value into the preset load change rate constraint equation for calculation;

[0039] The fused feature set of the training sample is input into the enterprise power load forecasting model, and the predicted enterprise power load value of the training sample is obtained through the input layer, RNN layer, self-attention layer and MLP layer in sequence;

[0040] Based on the defined total loss function, calculate the total loss caused by the predicted enterprise power load value, the actual enterprise power load value and the physical constraints of the current batch;

[0041] Apply the back-propagation algorithm to calculate the gradient of the total loss relative to the enterprise power load forecasting model, and iteratively update the parameters of the enterprise power load forecasting model;

[0042] Until the preset maximum number of iterations is reached, the training of the enterprise power load forecasting model is completed.

[0043] Optionally, the enterprise power load forecasting model specifically optimizes the model by minimizing both the forecast accuracy error and the operating cost, including:

[0044] Input the decision variable vector, call the trained enterprise power load forecasting model, predict the decision variable vector, and obtain the corresponding enterprise power load sequence;

[0045] Calculate the root mean square error between the enterprise power load sequence and the reference enterprise power load sequence, and define the root mean square error as the first objective function corresponding to the input decision variable vector;

[0046] Obtain predefined time-of-use electricity price data and define a cost accounting function. Through the cost accounting function, perform the following core calculations: multiply the load value of each time step of the enterprise's power load series by the corresponding time-of-use electricity price, and accumulate and sum the products of all time steps to obtain the total expected operating cost;

[0047] The total expected operating cost is defined as the second objective function corresponding to the input decision variable vector;

[0048] The first objective function and the second objective function are combined, and constraints are defined at the same time to form the objective function of the enterprise power load forecasting model.

[0049] Optionally, the constraints include: decision variable boundary constraints, total load capacity constraints, load ramp rate constraints and equipment operation constraints.

[0050] Optionally, the objective function is calculated as follows:

[0051]

[0052] Where Obj(X) is the objective function, w1 and w2 are non-negative weights, T is the total number of discrete time steps within the prediction time range, and L pred (X) is the enterprise power load sequence at the t-th time step predicted by the enterprise power load forecasting model under the condition of a given decision variable vector X, X is the decision variable vector, L ref_t is the reference enterprise power load sequence at the tth time step, Price t is the predefined time-of-use electricity price at the tth time step, Δt is the duration of each discrete time step, Cost add (X) is an additional cost item.

[0053] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0054] The present invention is based on a load forecasting and constrained multi-objective optimization method that integrates a physical information neural network with a self-attention hybrid model. On the one hand, by integrating physical mechanisms into a deep learning model, the PINN-SAHM can more deeply understand the inherent driving forces of load changes, significantly improving the ability to capture the nonlinear and time-varying characteristics of complex industrial loads and the prediction accuracy. It is particularly robust in scenarios with low data quality or sudden changes in operating conditions. On the other hand, combined with a constrained multi-objective particle swarm optimization algorithm, it can simultaneously optimize multiple conflicting objectives and strictly adhere to multiple practical constraints such as equipment operation, grid contract capacity, and load ramp rate during the optimization process. Furthermore, the present invention achieves a tight coupling of prediction and optimization, enabling forward-looking decision-making based on more accurate load forecast information, generating a series of Pareto-optimal energy-saving strategies covering different risk preferences and operational priorities for decision makers to choose from. By dynamically focusing on key influencing factors through the self-attention mechanism, it can better adapt to changes in the external environment, making the final energy-saving strategy more adaptable and effective, helping enterprises achieve deeper energy efficiency improvements and cost savings, and laying a solid foundation for participating in the future demand-side response market. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic flow chart of the power saving method based on enterprise power load forecasting provided by the present invention;

[0056] Figure 2This is a structural principle diagram of the enterprise power load forecasting model provided by the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the present invention more apparent, the following will provide a clear and complete description of the technical solutions of the present invention in conjunction with the accompanying drawings. It should be understood that the description is only a portion of the present invention, not all of it. The components of the present invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0058] Reference Figure 1 As shown, Figure 1 A flow chart of the power saving method based on enterprise power load forecasting provided by the present invention.

[0059] In some embodiments, a power saving method based on enterprise power load forecasting includes the following steps:

[0060] Obtain the company's historical power load time series data and historical power market data;

[0061] Perform feature extraction on the enterprise's historical power load time series data and the historical power market data respectively to obtain a first feature set corresponding to the enterprise's historical power load time series data and a second feature set corresponding to the historical power market data;

[0062] Combining the first feature set with the second feature set to generate a fused feature set;

[0063] Based on the PINN-SAHM neural network, an enterprise power load forecasting model is constructed. The fused feature set is input into the enterprise power load forecasting model for calculation until the training of the enterprise power load forecasting model is completed.

[0064] The fusion feature set obtained in real time is solved based on the trained enterprise power load forecasting model. The optimal power-saving strategy is solved according to the enterprise power load forecast value and the defined objective function.

[0065] In this embodiment, the data preparation phase is first performed, which involves obtaining the enterprise's historical power load time series data, such as hourly or higher frequency load records collected from the enterprise's internal energy management system, and obtaining relevant historical power market data, such as historical time-of-use electricity prices, peak and valley period divisions and other information. After obtaining the original data, pre-processing steps such as data cleaning, alignment and normalization are required. Entering the feature engineering phase, deep feature extraction is performed on the pre-processed enterprise historical power load time series data and historical power market data respectively. By applying time series analysis technology, statistical methods and domain knowledge, a first feature set (derived from the enterprise's historical power load time series data) and a second feature set (derived from historical power market data) that can effectively characterize the inherent laws of the load and the dynamics of market prices are generated. These two dimensionally rich feature sets are effectively combined, specifically by vector splicing, to generate a fusion feature set containing comprehensive information. This feature set will serve as the input basis for the subsequent prediction model. During the model construction and training phase, this embodiment constructs an enterprise power load forecasting model based on an advanced physical information neural network (PINN) fused with a self-attention hybrid model (SAHM) architecture. This model aims to combine the powerful fitting capabilities of data-driven learning with the constraints of prior knowledge of physical laws, and capture key influencing factors through a self-attention mechanism. A fused feature set generated from historical data is fed into the enterprise power load forecasting model for iterative calculations. Model parameters are adjusted using backpropagation and optimization algorithms, and the loss term of physical constraints is considered during training to minimize prediction error and ensure physical consistency. This completes the training of the enterprise power load forecasting model, resulting in a high-performance predictor with generalization capabilities. During the practical application phase, based on the trained enterprise power load forecasting model, the fused feature set representing the current and future states is solved in real time to obtain a sequence of accurate predicted values for future power load. Based on the enterprise power load forecast values, a constrained multi-objective particle swarm optimization algorithm (CMOPSO) is used to solve the problem, searching for and determining an equipment operation plan that optimally balances various objectives while satisfying all constraints. This ultimately serves as the optimal energy-saving strategy, which guides the enterprise's actual energy consumption behavior.

[0066] Specifically, the acquisition of the enterprise's historical power load time series data and historical power market data also includes a preprocessing process, which is specifically as follows:

[0067] Fill missing values and smooth outliers in the company's historical power load time series data;

[0068] Standardize historical electricity market data;

[0069] And align the time stamps of the enterprise's historical power load time series data with the corresponding historical power market data;

[0070] Complete data preprocessing of the company's historical power load time series data and historical power market data.

[0071] In a specific embodiment, this embodiment uses interpolation techniques, such as linear interpolation, spline interpolation, or the mean / median filling method based on adjacent data points, to effectively fill missing values in the data series to ensure data continuity. Furthermore, to eliminate potential interference from measurement errors or extreme events that may affect data analysis, statistical methods such as moving average or median filtering are used to smooth identified outliers, adjusting them to a reasonable range or replacing them with estimated values, thereby enhancing data stability and representativeness. Data normalization is performed on indicators of different dimensions in the historical electricity market data to eliminate differences in numerical scales between different features and prevent certain features from dominating due to excessively large values during model training, thereby accelerating model convergence and improving performance. Finally, to ensure consistency in the time bases of data from different sources, the processed enterprise historical electricity load time series data is strictly aligned with the corresponding historical electricity market data according to timestamps, generating a uniformly formatted and time-synchronized dataset. This provides reliable data support for the subsequent construction of a high-quality fusion feature set and training of the enterprise power load forecasting model.

[0072] In the specific implementation of this embodiment, the feature extraction is performed on the enterprise's historical power load time series data and the historical power market data respectively. The feature extraction is performed on the enterprise's historical power load time series data, and the specific steps are as follows:

[0073] Applying a predetermined wavelet mother function to the enterprise's historical power load time series data to perform multi-layer wavelet packet decomposition to obtain multiple wavelet packet node signals covering different frequency bands;

[0074] Calculating a characteristic quantity of at least a portion of the wavelet packet node signal, wherein the characteristic quantity is selected from at least one of energy, entropy, variance, skewness, or kurtosis;

[0075] The calculated feature quantities are combined into the first feature set.

[0076] The specific steps for feature extraction of historical electricity market data are as follows:

[0077] Identify key market events of predefined types and their occurrence times in historical electricity market data;

[0078] For the identified key market events, extract the historical power load subsequence within the corresponding time window;

[0079] The historical power load subsequences are processed based on the symbolic aggregation approximation algorithm to generate a symbol string representing the load pattern;

[0080] Mining the frequent association patterns between key market events and generated load pattern symbol strings;

[0081] The second feature set is constructed based on the mined frequent association patterns.

[0082] representing the first feature set as a first numerical feature vector;

[0083] representing the second feature set as a second numerical feature vector;

[0084] Sequentially appending the second numerical feature vector to the first numerical feature vector to form a single fused feature vector whose dimension is the sum of the dimensions of the first and second numerical feature vectors;

[0085] The single fused feature vector is determined as the fused feature set, and the fused feature set is input into the enterprise power load forecasting model.

[0086] In the implementation, this embodiment further performs feature engineering and fusion, and performs deep feature extraction on the pre-processed enterprise historical power load time series data and historical power market data to generate the first feature set and the second feature set. hist ={l1,l2,...,l N}, where l i is the load value at the i-th moment, and N is the sequence length. When extracting features, a predetermined wavelet mother function is applied to perform J-layer multi-layer wavelet packet decomposition to obtain multiple wavelet packet node signals S covering different frequency bands. J,k , where k is the node index: Calculate wavelet packet node signal The characteristic quantity of J,k,1 is the signal coefficient, P k is the signal length. The characteristic quantity is selected from: energy To reflect the signal strength; entropy in, To measure signal complexity; variance Among them, μ J,k is the signal mean, which is used to measure the degree of signal fluctuation; skewness To measure the asymmetry of the distribution; kurtosis All the calculated selected feature quantities are combined in a predetermined order to form the first feature set. At the same time, the historical electricity market data M histTo perform feature extraction, first identify key market events of predefined types and their occurrence times in historical electricity market data to obtain an event list. For each identified key market event, extract the corresponding historical power load subsequence within a predetermined time window near its occurrence time. Based on the symbolic aggregation approximation (SAX) algorithm, set the word length and alphabet size, process each historical power load subsequence, generate a symbol string representing the load pattern, and use the formed event-pattern pair data set to mine the frequent association patterns between the key market events and the generated load pattern symbol string through frequent pattern mining technology. Based on the mined frequent association patterns, construct a second feature set. After the feature extraction is completed, the second numerical feature vector represented by the second feature set is sequentially appended to the first numerical feature vector represented by the first feature set through vector concatenation operations to form a single fused feature vector of one dimension, which is represented as a fused feature set. This fused feature set is then input into the subsequent enterprise power load forecasting model for training or prediction.

[0087] like Figure 2 As shown, in the specific application of this embodiment, the enterprise power load forecasting model is constructed based on the PINN-SAHM neural network, which is specifically:

[0088] An input layer of the enterprise power load forecasting model is defined and configured to receive the fused feature set as input;

[0089] An RNN layer is used to process the sequence output by the input layer to capture the temporal dependency in the fused feature set and output an RNN output sequence containing the hidden state of each time step;

[0090] a self-attention mechanism layer, configured to receive the RNN output sequence as input, calculate attention weights between hidden states at different time steps in the RNN output sequence, and perform a weighted summation of the RNN output sequence based on the attention weights to generate a context vector;

[0091] The MLP layer receives the context vector as input, performs nonlinear transformation through at least one fully connected layer, and outputs the enterprise power load forecast value.

[0092] The training process of the enterprise power load forecasting model is as follows:

[0093] Construct a training data set, where each training sample contains a fusion feature set corresponding to a historical time period and the corresponding actual enterprise power load value;

[0094] Define the total loss function of the enterprise power load forecasting model. The total loss function consists of a data-driven loss term and a physical information constraint loss term, and is weighted and summed by a preset weight λ. The data-driven loss term is specifically the mean square error between the predicted enterprise power load value output by the enterprise power load forecasting model and the corresponding actual enterprise power load value in the training sample; the physical information constraint loss term is specifically the mean square value of the residual of the load change rate constraint equation obtained by substituting the predicted enterprise power load value into the preset load change rate constraint equation for calculation;

[0095] The fused feature set of the training sample is input into the enterprise power load forecasting model, and the predicted enterprise power load value of the training sample is obtained through the input layer, RNN layer, self-attention layer and MLP layer in sequence;

[0096] Based on the defined total loss function, calculate the total loss caused by the predicted enterprise power load value, the actual enterprise power load value and the physical constraints of the current batch;

[0097] Apply the back-propagation algorithm to calculate the gradient of the total loss relative to the enterprise power load forecasting model, and iteratively update the parameters of the enterprise power load forecasting model;

[0098] Until the preset maximum number of iterations is reached, the training of the enterprise power load forecasting model is completed.

[0099] In specific implementation, this embodiment defines the architecture of an enterprise power load forecasting model. Its input layer is configured to receive as input the fused feature set generated in the preceding steps. Following the input layer is a recurrent neural network (RNN) layer, which processes the temporal information output by the input layer, effectively capturing the implicit temporal dependencies within the fused feature set and outputting a sequence of RNN outputs containing hidden states at each time step. A self-attention mechanism layer is provided, which receives as input the hidden state sequence output by the RNN. The self-attention mechanism layer dynamically assesses the contribution of each historical state to the current forecast by calculating the mutual importance (attention weights) between hidden states at different time steps in the sequence. Based on these importances, it performs a weighted integration of the RNN output sequence, ultimately generating one or more context vectors that condense key temporal information. Finally, the model's output consists of a multi-layer perceptron (MLP) layer, which receives as input the context vector generated by the self-attention layer. This layer undergoes nonlinear transformation and feature mapping through at least one fully connected layer containing a nonlinear activation function, ultimately outputting a specific enterprise power load forecast value. Based on the above structure, in this embodiment, during the training iteration, the fused feature set of the training samples is input into the enterprise power load forecasting model, and the data passes through the input layer, RNN layer, self-attention layer and MLP layer in sequence to obtain the predicted enterprise power load value of the batch of samples. Based on the defined total loss function, the predicted value, actual value and total error caused by the physical constraints of the current batch are calculated. The back propagation algorithm is applied to calculate the gradient of the total loss relative to all adjustable parameters of the enterprise power load forecasting model: weights and biases. Using these gradients, the parameters of the enterprise power load forecasting model are iteratively adjusted through optimization algorithms such as Adam to gradually reduce the total loss. This process continues until the training reaches the preset maximum number of iterations, and the training of the enterprise power load forecasting model is completed.

[0100] In further implementation, the enterprise power load forecasting model specifically aims to minimize both the forecast accuracy error and the operating cost, including:

[0101] Input the decision variable vector, call the trained enterprise power load forecasting model, predict the decision variable vector, and obtain the corresponding enterprise power load sequence;

[0102] Calculate the root mean square error between the enterprise power load sequence and the reference enterprise power load sequence, and define the root mean square error as the first objective function corresponding to the input decision variable vector;

[0103] Obtain predefined time-of-use electricity price data and define a cost accounting function. Through the cost accounting function, perform the following core calculations: multiply the load value of each time step of the enterprise's power load series by the corresponding time-of-use electricity price, and accumulate and sum the products of all time steps to obtain the total expected operating cost;

[0104] The total expected operating cost is defined as the second objective function corresponding to the input decision variable vector;

[0105] The first objective function is combined with the second objective function, and constraints are defined including: decision variable boundary constraints, total load capacity constraints, load ramp rate constraints and equipment operation constraints, to form the objective function of the enterprise power load forecasting model.

[0106] In the specific implementation of this embodiment, this embodiment further applies the previously trained self-attention hybrid enterprise power load forecasting model (PINN-SAHM) that integrates physical information to guide the enterprise's operational decisions within an optimization framework that aims to simultaneously minimize the deviation between the predicted load and the target load and the expected operating power cost. The specific implementation of this process is as follows:

[0107] Define the input decision variable vector X of the optimization problem. In this embodiment, it is assumed that the key factors that can be controlled by the enterprise are the start-up and shutdown plans of the main production lines and the charging and discharging strategy of the energy storage system. Therefore, the decision variable vector X specifically includes: (a) the planned start-up time t of each major high-energy-consuming production line p start,p and operating mode / power level m p (b) The planned charging power P of the enterprise's own energy storage system (if equipped) at each time step t in the future optimization time domain chg,t and discharge power P dis,t For a given instance of a decision variable vector X, i.e., a set of specific production start-up times and energy storage operation plans, the previously trained enterprise power load forecasting model is called. In order to obtain the load forecast under this decision X, the following steps are performed: based on the forecast of future uncontrollable factors such as weather, base load and expected market events, a basic future feature sequence is generated by referring to the historical feature extraction method; the operation represented by the decision variable X, such as production line p at t start,p The load increases due to startup, energy storage charging P chg,t The load increases, the discharge P dis,tThe resulting load reduction is superimposed on the base load pattern. Future scenario features incorporating decision information are input into the trained enterprise power load forecasting model to obtain the corresponding enterprise power load forecast series for the future period. The key metrics for evaluating the effectiveness of decision X are then calculated. First, the root mean square error (RMS) between the predicted enterprise power load series and a preset reference enterprise power load series is calculated. As can be understood, the reference enterprise power load series represents an ideal stable operating curve, a curve that meets specific production targets, or a grid-friendly interaction curve. The RMS error is defined as the first objective function corresponding to the input decision variable vector X. The goal is to make actual operation as close to the reference state as possible, that is, to minimize the first objective function. Second, the economic cost of the decision is evaluated. Predefined time-of-use (TOU) price data is obtained. This data provides the unit price of electricity for each time step t within the optimization domain. For each time step t in the predicted enterprise power load series, the predicted load value is multiplied by the corresponding TOU price to obtain the electricity cost for that time step. The costs for all time steps are summed to obtain the total expected operating power cost under decision X. This total cost is defined as the second objective function corresponding to the input decision variable vector X. The goal is to minimize the company's electricity expenditure, that is, to minimize the second objective function. The first objective function and the second objective function, which are intended to be minimized, are combined to form the final goal of the optimization problem. Specifically, non-negative weights are set to form a weighted single objective function to find the Pareto optimal solution set. At the same time, the optimization process must meet the defined constraints, which specifically include: Decision variable boundary constraints: Ensure that the production line startup time t start,p During the allowed working hours, the operation mode m p It is a valid option, the energy storage charging and discharging power P chg,t ,P dis,t It does not exceed its rated power and is non-negative, while considering that the state of charge of the energy storage system must be maintained within a safe operating range. Total load capacity constraint: It requires that at any time step t, the predicted total load must not exceed the maximum demand contract value signed between the enterprise and the power grid or the upper limit of the transformer capacity. Load ramp rate constraint: The absolute value of the change rate of the predicted load between adjacent time steps is limited to not exceed the maximum allowed value to protect the equipment and meet the requirements of the power grid. Equipment operation constraints: For example, a specific production line p may have minimum continuous operating time and minimum downtime requirements; the energy storage system may have a maximum daily charge and discharge times or total power limit, etc. By solving the combined enterprise power load forecasting model, dual optimization objectives and constraints, a set of optimal decision variables, namely the optimal production start-up time and energy storage charge and discharge plan, can be obtained, thereby guiding the enterprise to achieve the best balance between forecast accuracy and operating costs while meeting production and operation requirements.

[0108] The calculation formula of the objective function is:

[0109]

[0110] Where Obj(X) is the objective function, w1 and w2 are non-negative weights, T is the total number of discrete time steps within the prediction time range, and L pred (X) is the enterprise power load sequence at the t-th time step predicted by the enterprise power load forecasting model under the condition of a given decision variable vector X, X is the decision variable vector, L ref_t is the reference enterprise power load sequence at the tth time step, Price t is the predefined time-of-use electricity price at the tth time step, Δt is the duration of each discrete time step, Cost add (X) is an additional cost item.

[0111] The above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A power saving method based on enterprise power load forecasting, characterized in that: The steps of the method include: Obtain the company's historical power load time series data and historical power market data; Perform feature extraction on the enterprise's historical power load time series data and the historical power market data respectively to obtain a first feature set corresponding to the enterprise's historical power load time series data and a second feature set corresponding to the historical power market data; Combining the first feature set with the second feature set to generate a fused feature set; Build an enterprise power load forecasting model based on the PINN-SAHM neural network, input the fused feature set into the enterprise power load forecasting model for calculation until the training of the enterprise power load forecasting model is completed; The fusion feature set obtained in real time is solved based on the trained enterprise power load forecasting model. The optimal power-saving strategy is solved according to the enterprise power load forecast value and the defined objective function.

2. The power saving method based on enterprise power load forecasting according to claim 1 is characterized in that: The acquisition of the enterprise's historical power load time series data and historical power market data also includes a preprocessing process, which is specifically as follows: Fill missing values and smooth outliers in the company's historical power load time series data; Standardize historical electricity market data; And align the time stamps of the enterprise's historical power load time series data with the corresponding historical power market data; Complete data preprocessing of the company's historical power load time series data and historical power market data.

3. The power saving method based on enterprise power load forecasting according to claim 2, characterized in that: The feature extraction is performed on the enterprise's historical power load time series data and the historical power market data respectively, wherein the feature extraction is performed on the enterprise's historical power load time series data, and the specific steps are: Applying a predetermined wavelet mother function to the enterprise's historical power load time series data to perform multi-layer wavelet packet decomposition to obtain multiple wavelet packet node signals covering different frequency bands; Calculating a characteristic quantity of at least a portion of the wavelet packet node signal, wherein the characteristic quantity is selected from at least one of energy, entropy, variance, skewness, or kurtosis; The calculated feature quantities are combined into the first feature set.

4. The power saving method based on enterprise power load forecasting according to claim 3 is characterized in that: The specific steps for feature extraction of historical electricity market data are as follows: Identify key market events of predefined types and their occurrence times in historical electricity market data; For the identified key market events, extract the historical power load subsequence within the corresponding time window; The historical power load subsequences are processed based on the symbolic aggregation approximation algorithm to generate a symbol string representing the load pattern; Mining the frequent association patterns between key market events and generated load pattern symbol strings; The second feature set is constructed based on the mined frequent association patterns.

5. The power saving method based on enterprise power load forecasting according to claim 4 is characterized in that: The combination of the first feature set and the second feature set is specifically as follows: representing the first feature set as a first numerical feature vector; representing the second feature set as a second numerical feature vector; Sequentially appending the second numerical feature vector to the first numerical feature vector to form a single fused feature vector whose dimension is the sum of the dimensions of the first and second numerical feature vectors; The single fused feature vector is determined as the fused feature set, and the fused feature set is input into the enterprise power load forecasting model.

6. The power saving method based on enterprise power load forecasting according to claim 5, characterized in that: The enterprise power load forecasting model based on the PINN-SAHM neural network is specifically constructed as follows: An input layer of the enterprise power load forecasting model is defined and configured to receive the fused feature set as input; An RNN layer is used to process the sequence output by the input layer to capture the temporal dependency in the fused feature set and output an RNN output sequence containing the hidden state of each time step; a self-attention mechanism layer, configured to receive the RNN output sequence as input, calculate attention weights between hidden states at different time steps in the RNN output sequence, and perform a weighted summation of the RNN output sequence based on the attention weights to generate a context vector; The MLP layer receives the context vector as input, performs nonlinear transformation through at least one fully connected layer, and outputs the enterprise power load forecast value.

7. The power saving method based on enterprise power load forecasting according to claim 6, characterized in that: The training process of the enterprise power load forecasting model is as follows: Construct a training data set, where each training sample contains a fusion feature set corresponding to a historical time period and the corresponding actual enterprise power load value; Define the total loss function of the enterprise power load forecasting model. The total loss function consists of a data-driven loss term and a physical information constraint loss term, and is weighted and summed by a preset weight λ. The data-driven loss term is specifically the mean square error between the predicted enterprise power load value output by the enterprise power load forecasting model and the corresponding actual enterprise power load value in the training sample; the physical information constraint loss term is specifically the mean square value of the residual of the load change rate constraint equation obtained by substituting the predicted enterprise power load value into the preset load change rate constraint equation for calculation; The fused feature set of the training sample is input into the enterprise power load forecasting model, and the predicted enterprise power load value of the training sample is obtained through the input layer, RNN layer, self-attention layer and MLP layer in sequence; Based on the defined total loss function, calculate the total loss caused by the predicted enterprise power load value, the actual enterprise power load value and the physical constraints of the current batch; Apply the back-propagation algorithm to calculate the gradient of the total loss relative to the enterprise power load forecasting model, and iteratively update the parameters of the enterprise power load forecasting model; Until the preset maximum number of iterations is reached, the training of the enterprise power load forecasting model is completed.

8. The power saving method based on enterprise power load forecasting according to claim 7, characterized in that: The enterprise power load forecasting model specifically aims to minimize both the forecast accuracy error and the operating cost, and includes: Input the decision variable vector, call the trained enterprise power load forecasting model, predict the decision variable vector, and obtain the corresponding enterprise power load sequence; Calculate the root mean square error between the enterprise power load sequence and the reference enterprise power load sequence, and define the root mean square error as the first objective function corresponding to the input decision variable vector; Obtain predefined time-of-use electricity price data and define a cost accounting function. Through the cost accounting function, perform the following core calculations: multiply the load value of each time step of the enterprise's power load series by the corresponding time-of-use electricity price, and accumulate and sum the products of all time steps to obtain the total expected operating cost; The total expected operating cost is defined as the second objective function corresponding to the input decision variable vector; The first objective function and the second objective function are combined, and constraints are defined at the same time to form the objective function of the enterprise power load forecasting model.

9. The power saving method based on enterprise power load forecasting according to claim 8, characterized in that: The constraints include: decision variable boundary constraints, total load capacity constraints, load ramp rate constraints and equipment operation constraints.

10. The power saving method based on enterprise power load forecasting according to claim 9, characterized in that: The calculation formula of the objective function is: Where Obj(X) is the objective function, w1 and w2 are non-negative weights, T is the total number of discrete time steps within the prediction time range, and L pred (X) is the enterprise power load sequence at the t-th time step predicted by the enterprise power load forecasting model under the condition of a given decision variable vector X, X is the decision variable vector, L ref_t is the reference enterprise power load sequence at the tth time step, Price t is the predefined time-of-use electricity price at the tth time step, Δt is the duration of each discrete time step, Cost add (X) is an additional cost item.

Images