Combined heat and power generation load prediction method based on artificial intelligence

Abstract

The invention discloses a combined heat and power generation load prediction method based on artificial intelligence, and relates to the field of data analysis, and the method comprises the following steps: collecting historical operation data of a combined heat and power generation system, constructing a multi-source time series data set, and carrying out the cleaning, alignment and standardization processing of the data; time sequence features, statistical features and physical features based on domain knowledge are extracted from the multi-source time sequence data, and feature vectors are constructed; constructing a prediction model fusing the space-time diagram convolutional network and the multi-head time sequence transformer; historical data is used for training the hybrid prediction model, model parameters are optimized through a joint loss function, and the joint loss function comprises prediction error loss and physical consistency loss. According to the method, through the hybrid model of the space-time diagram convolution and the time sequence transformer, the space-time dynamic characteristics of the load data can be captured more comprehensively, and the prediction precision is obviously superior to that of a traditional method.

Description

Technical Field

[0001] This invention relates to the field of data analysis, specifically to a method for predicting combined heat and power loads based on artificial intelligence. Background Technology

[0002] Combined heat and power (CHP) is an efficient energy utilization method, but its "heat-driven power generation" or "electricity-driven heat generation" operating mode results in a high degree of coupling between electricity and heat loads, posing a significant challenge to the optimal scheduling of the system. Accurate load forecasting is crucial for achieving the economical and safe operation of CHP systems.

[0003] Existing load forecasting methods are mainly divided into two categories: traditional statistical methods (such as time series analysis and regression analysis) and shallow machine learning methods (such as support vector machines and gradient boosting trees). These methods have obvious limitations: First, they usually forecast electricity and heat loads separately, ignoring the strong coupling mechanism between the two, and the forecast results often violate the physical operating constraints of the system; second, these models are difficult to effectively handle multi-source heterogeneous data such as meteorological, calendar, and equipment status data, and have insufficient ability to capture long-term complex time dependencies.

[0004] Therefore, there is an urgent need for a new prediction method that can deeply integrate system physical characteristics and data-driven capabilities to achieve high-precision and physically consistent joint prediction of the electrical and thermal loads of cogeneration systems. Summary of the Invention

[0005] The purpose of this invention is to provide a solution to the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting combined heat and power (CHP) load based on artificial intelligence includes the following steps: S1. Data Acquisition and Preprocessing: Collect historical operating data of the cogeneration system, construct a multi-source time-series dataset, and perform data cleaning, alignment, and standardization. S2. Multimodal feature engineering: Extracting temporal features, statistical features, and domain-knowledge-based physical features from the multi-source time-series data to construct feature vectors; S3. Hybrid AI Model Construction: Construct a prediction model that integrates a spatiotemporal graph convolutional network and a multi-head time transformer. The model includes a graph convolutional module for capturing spatial dependencies, a time transformer module for capturing complex temporal dependencies, and a physical information embedding module that uses the system's physical equations as constraints. S4. Model Training and Optimization: The hybrid prediction model is trained using historical data, and the model parameters are optimized through a joint loss function, which includes prediction error loss and physical consistency loss. S5. Load Forecasting: Input real-time data into the trained hybrid forecasting model to obtain the joint forecast results of electrical load and heat load for a specific future time window.

[0007] As a further preferred embodiment of the present invention, in step S1, the historical operating data includes: historical electrical load values, historical heat load values, meteorological data, calendar type data, and key operating parameters of the combined heat and power unit.

[0008] As a further preferred embodiment of the present invention, in step S2, the physical characteristics based on domain knowledge include: the rate of change of the thermoelectric ratio, the estimated value of the system's overall energy efficiency, and the thermal inertia index based on the characteristics of the heating network.

[0009] As a further preferred embodiment of the present invention, in step S3, the graph structure of the spatiotemporal graph convolutional network is constructed according to the physical layout of the cogeneration system, the nodes represent key equipment, and the edge weights are determined by the physical connection relationship or load correlation between the equipment.

[0010] As a further preferred embodiment of the present invention, in step S3, the physical information embedding module calculates the physical consistency loss by substituting the intermediate prediction results of the model into the simplified system energy balance equation, and backpropagates it to constrain the training process of the model.

[0011] As a further preferred embodiment of the present invention, in step S4, the joint loss function is defined as: L=αLpred+βLphysics Where Lpred is the mean squared error between the predicted and actual values, Lphysics is the physical consistency loss, and α and β are hyperparameters used to balance the two losses.

[0012] As a further preferred embodiment of the present invention, the method can also perform uncertainty quantification by introducing quantile regression or Monte Carlo Dropout technology into the model output layer to output the probability distribution range of the load prediction value.

[0013] Compared with the prior art, the beneficial effects of the present invention are: Higher accuracy: By using a hybrid model of spatiotemporal graph convolution and time-series transformer, the spatiotemporal dynamic characteristics of load data can be captured more comprehensively, and the prediction accuracy is significantly better than that of traditional methods.

[0014] Physical consistency: By introducing physical information constraints, we ensure that the AI's predictions conform to basic physical laws such as thermodynamics and energy conservation, thus avoiding absurd predictions and enhancing the model's reliability and generalization ability.

[0015] Strong generalization ability: The model can learn the inherent operating rules of the system rather than the superficial data correlations, thus showing stronger robustness when facing new operating conditions or climate change. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In this embodiment of the invention, a method for predicting cogeneration load based on artificial intelligence is described in detail below: S1. Data Acquisition and Preprocessing: First, historical operating data is collected from the cogeneration system, specifically including: historical electrical load values ​​(kW) and heat load values ​​(GJ / h) sampled at 15-minute intervals; meteorological data (ambient temperature, humidity, precipitation, wind speed, etc.); calendar-type data (weekday / holiday identifiers, seasonal information); and key operating parameters of the cogeneration unit (turbine inlet steam pressure, boiler outlet temperature, heating network circulation flow, etc.). Next, a multi-source time-series dataset is constructed, and the following preprocessing operations are performed: data cleaning (outliers are removed using the 3σ criterion, and missing values ​​are filled using linear interpolation); time axis alignment (the sampling frequency is standardized to 15 minutes / point); and standardization (the Z-score method is used to transform the feature data to a distribution range with a mean of 0 and a standard deviation of 1) to ensure data quality and consistency.

[0018] S2. Multimodal feature engineering: From the preprocessed multi-source time series data, three types of features are extracted to construct feature vectors: (1) Time series features: The trend term (first-order difference), periodic term (daily / weekly periodic components extracted by Fourier transform), and autocorrelation coefficient (lag of 1-24 hours) of the load sequence are calculated by the sliding window method; (2) Statistical features: The mean, variance, kurtosis, skewness, and maximum / minimum difference within the 24-hour sliding window are calculated; (3) Physical features: The heat-to-electricity ratio change rate (ΔQ / ΔP), system comprehensive energy efficiency (total output energy / fuel consumption), and heat network thermal inertia index (pipeline heat storage coefficient × transmission distance) are derived based on thermodynamic principles. The feature vector dimension is controlled within 128 dimensions, and multicollinearity features (VIF<10) are removed by using variance expansion factor (VIF).

[0019] S3. Hybrid Artificial Intelligence Model Construction: Construct a three-layer cascaded prediction model: (1) Graph Convolution Module: Adopting the improved ChebNet architecture, the topology of the cogeneration system (nodes include 12 types of equipment such as boilers, steam turbines, and heat exchangers) is used as input, and the steam / hot water transmission dependency between equipment is captured through two layers of graph convolution (hidden dimension 64→32); (2) Time Transformer Module: A four-layer multi-head attention network (8 heads of attention, hidden dimension 128) is adopted, combined with position encoding and time embedding, to model the long-term dependency of the load sequence (capturing the time sequence pattern of up to 7 days × 96 points / day); (3) Physical Information Embedding Module: An energy conservation constraint layer is inserted in the middle layer of the model, and the predicted electric / heat load is substituted into the simplified energy balance equation (input fuel energy = electric output + heat output + system loss), and the physical deviation is calculated as a regularization term. The model output layer adopts a dual-branch structure, outputting the predicted values ​​of electric load (MLP layer + linear activation) and heat load (MLP layer + ReLU activation) respectively.

[0020] S4. Model Training and Optimization: The following process was used for model training: (1) Dataset partitioning: The dataset was divided into a training set (data from 2019-2021), a validation set (data from 2022), and a test set (data from 2023) in a ratio of 7:2:1; (2) Training parameter settings: AdamW optimizer (initial learning rate 0.001, cosine annealing scheduling), batch size 64, 200 training rounds, early stopping strategy (stop if there is no improvement in the validation set loss after 50 rounds); (3) Joint loss function optimization: L=0.7Lpred+0.3Lphysics, where Lpred is the weighted sum of mean squared errors (MSE) of electrical / thermal loads (weight ratio 1:1.5), and Lphysics is the L1 loss of the energy balance equation bias. The hyperparameters α=0.7 and β=0.3 were determined by grid search. The final model had an electrical load MAE ≤2.5% and a thermal load MAE ≤3.2% on the test set.

[0021] S5. Load Forecasting and Uncertainty Quantification: The real-time forecasting process includes: (1) Online Inference: Inputting the 15-minute data collected in real time (preprocessed by S1) into the trained model, and outputting the joint forecast sequence of electricity / heat load for the next 24 hours (96 time points); (2) Uncertainty Quantification: Using the Monte Carlo Dropout method (maintaining a 5% dropout rate during the inference phase), 50 random forward propagations are performed to generate the probability distribution interval of the predicted value (confidence level 95%); (3) Result Calibration: Adjusting the probability interval through quantile regression to ensure that the frequency of the actual value falling within the prediction interval is ≥90%. TensorRT is used to accelerate inference during system deployment, and the single sample prediction time is ≤50ms, which meets the real-time requirements.

[0022] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0023] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting combined heat and power load based on artificial intelligence, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Collect historical operating data of the cogeneration system, construct a multi-source time-series dataset, and perform data cleaning, alignment, and standardization. S2. Multimodal feature engineering: Extracting temporal features, statistical features, and domain-knowledge-based physical features from the multi-source time-series data to construct feature vectors; S3. Hybrid AI Model Construction: Construct a prediction model that integrates a spatiotemporal graph convolutional network and a multi-head time transformer. The model includes a graph convolutional module for capturing spatial dependencies, a time transformer module for capturing complex temporal dependencies, and a physical information embedding module that uses the system's physical equations as constraints. S4. Model Training and Optimization: The hybrid prediction model is trained using historical data, and the model parameters are optimized through a joint loss function, which includes prediction error loss and physical consistency loss. S5. Load Forecasting: Input real-time data into the trained hybrid forecasting model to obtain the joint forecast results of electrical load and heat load for a specific future time window.

2. The method according to claim 1, characterized in that, In step S1, the historical operating data includes: historical electrical load values, historical heat load values, meteorological data, calendar-type data, and key operating parameters of the combined heat and power unit.

3. The method according to claim 1, characterized in that, In step S2, the physical characteristics based on domain knowledge include: the rate of change of the thermoelectric ratio, the estimated value of the system's overall energy efficiency, and the thermal inertia index based on the characteristics of the heating network.

4. The method according to claim 1, characterized in that, In step S3, the graph structure of the spatiotemporal graph convolutional network is constructed based on the physical layout of the cogeneration system. Nodes represent key equipment, and edge weights are determined by the physical connection relationship or load correlation between equipment.

5. The method according to claim 1, characterized in that, In step S3, the physical information embedding module calculates the physical consistency loss by substituting the intermediate prediction results of the model into the simplified system energy balance equation, and backpropagates it to constrain the training process of the model.

6. The method according to claim 1, characterized in that, In step S4, the joint loss function is defined as: L=αLpred+βLphysics Where Lpred is the mean squared error between the predicted and actual values, Lphysics is the physical consistency loss, and α and β are hyperparameters used to balance the two losses.

7. The method according to claim 1, characterized in that, The method can also quantify uncertainty by introducing quantile regression or Monte Carlo Dropout techniques into the model output layer to output the probability distribution range of the load forecast.