Mine equipment energy consumption prediction method

Abstract

The invention discloses a mining equipment energy consumption prediction method, and relates to the technical field of mining equipment energy consumption prediction.The mining equipment energy consumption prediction method comprises the steps that through multi-dimensional data collection, operation, process and environment parameters are collected through a sensor cluster; noise reduction is carried out through a generative adversarial network in combination with empirical mode decomposition, and data is restored through a space-time interpolation network; identifying working conditions and extracting features by means of a hidden Markov model in combination with an attention mechanism; a cross-device transfer learning framework is constructed, and a cloud training general model is combined with local data fine tuning; the edge end deploys a lightweight model for real-time prediction, and the cloud end generates a global energy-saving strategy; through digital twinborn visualization, a model and a strategy are automatically corrected based on SHAP value analysis. The equipment idling rate is reduced; the unit energy consumption of the crushing link is reduced; the abnormal response time is shortened; and the prediction precision and the system adaptability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of energy consumption prediction of mining equipment, and in particular to a method for predicting energy consumption of mining equipment. Background Art

[0002] As the mining industry transitions toward intelligent and green operations, energy consumption prediction technology for mining equipment has become crucial for improving energy efficiency and reducing operating costs. Traditional energy consumption prediction methods primarily rely on single-dimensional operating data, such as current and speed, combined with linear regression or shallow machine learning models for prediction. However, the complex and ever-changing mining production environment, with equipment operating conditions dynamically switching between excavation, transportation, and no-load conditions, and the influence of multiple factors such as ore hardness and environmental parameters, present challenges for existing technologies: Insufficient data fusion: Mining equipment energy consumption involves heterogeneous data from multiple sources, including operating parameters, production processes, and the environment. Existing systems often rely on independent data collection and lack spatiotemporal correlation modeling, resulting in the loss of key features and an inability to fully reflect the drivers of energy consumption.

[0003] Weak model generalization: Traditional models rely on historical data from specific equipment for training, making them difficult to adapt to the differences between different equipment models or mining scenarios. For example, the energy consumption characteristics of crushers and dump trucks vary significantly, but existing transfer learning technologies fail to effectively extract common energy consumption patterns across devices, resulting in a sharp drop in model prediction accuracy in new scenarios.

[0004] Poor real-time performance: Mining equipment must respond to changes in operating conditions in real time to achieve energy-saving control. However, traditional methods rely on centralized cloud-based computing, which results in high data transmission latency and limited edge computing capabilities, making it difficult to meet the needs of millisecond-level real-time predictions. Furthermore, inefficiencies in missing data repair and noise filtering further impact the timeliness of predictions.

[0005] Furthermore, existing data preprocessing methods often employ wavelet transforms or Kalman filters for noise reduction, but these methods lack robustness to non-stationary noise. Missing data repair often relies on time series interpolation, ignoring the spatial correlations between devices. Furthermore, static classification methods based on threshold rules or support vector machines struggle to capture the temporal dependencies of dynamic switching between operating conditions, leading to delayed recognition or misjudgment.

[0006] These problems have resulted in the prediction accuracy of traditional methods generally being below 80%, the equipment idling rate being high, exceeding 30%, and the space for optimizing unit energy consumption being limited, making it impossible to meet the efficient energy-saving needs of intelligent mines. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for predicting energy consumption of mining equipment.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a method for predicting energy consumption of mining equipment, comprising the following steps: Multi-dimensional data collection: A sensor cluster collects real-time operating data, production process data, and environmental parameters of mining equipment. The operating data includes equipment vibration, temperature, current, rotation speed, and feed rate. The production process data includes blasting parameters, ore grade, and ore size distribution. The environmental parameters include humidity, air pressure, and ore hardness. Data preprocessing: De-noise and repair missing data from the collected raw data. Generative adversarial networks combined with empirical mode decomposition are used to separate valid signals from noise. Missing data is reconstructed based on the correlation between the spatial location of the device and the time series using a spatiotemporal interpolation network. Dynamic operating condition identification and feature extraction: Hidden Markov models are used to identify equipment operating conditions in real time, including excavation, transportation, no-load, full-load, half-load, and standby. Key energy consumption characteristics under different operating conditions are extracted using an attention mechanism. Adaptive prediction model construction: Build a cross-device transfer learning framework to train a universal energy consumption prediction model in the cloud using historical data from multiple mines and multiple equipment models. Then, use a small amount of local data from the target equipment to fine-tune the parameters of the universal model. Edge-cloud collaborative prediction and optimization: A lightweight prediction model is deployed on the device side for real-time energy consumption prediction and abnormal warning. In the cloud, a model predictive control algorithm is used based on data from the entire mine area to generate a global energy-saving strategy. The lightweight model compresses the parameters of the general model through knowledge distillation. Prediction-control closed-loop feedback: Map prediction results to device control parameters, visualize energy consumption status and strategy execution effects through digital twins, analyze model decision-making based on SHAP values, and automatically correct prediction models and control strategies.

[0009] As a preferred technical solution of the present invention, the sensor cluster includes low-cost MEMS sensors and ore hardness testers, which realize the synchronous collection of operation data, production process data and environmental parameters and the construction of a three-level data grid through the Internet of Things. The three-level data grid covers single equipment points, regions and the entire mining area.

[0010] As a preferred technical solution of the present invention, in the data preprocessing step, the generative adversarial network includes a noise generator and a signal discriminator. Through adversarial training, the signal discriminator distinguishes between real and effective signals and pseudo-noise signals generated by the noise generator, and empirical mode decomposition performs multi-scale decomposition on the real and effective signals to extract intrinsic mode functions.

[0011] As an optimal technical solution of the present invention, the spatiotemporal interpolation network constructs a device spatial position association matrix based on a graph neural network, and combines it with a long short-term memory network to process time series data, and realizes the reconstruction of missing data through spatiotemporal joint modeling.

[0012] As a preferred technical solution of the present invention, in the dynamic working condition identification and feature extraction steps, the observation sequence of the hidden Markov model is the time series feature vector of the equipment operation data, and the transition probability of each working condition is calculated by the forward-backward algorithm. The attention mechanism weights the energy consumption characteristics under different working conditions to highlight the key influencing factors.

[0013] As a preferred technical solution of the present invention, the cross-device transfer learning framework includes: A general feature extraction layer based on a deep neural network, trained with multi-source device data to learn common energy consumption characteristics across devices; The output layer specific to the target device fine-tunes the parameters of the general feature extraction layer using the local data of the target device to form a prediction model adapted to the specific device.

[0014] As a preferred technical solution of the present invention, in the edge-cloud collaboration step, the parameter scale of the lightweight prediction model is compressed by 60%-80% compared with the original deep learning model, supports localized prediction under offline conditions, and has a continuous running time of ≥72 hours when the network is disconnected.

[0015] As a preferred technical solution of the present invention, the global energy-saving strategy includes: Real-time planning of the optimal driving path for dump trucks based on ore grade and transportation routes; Adjust the start and stop timing of high-power equipment based on real-time electricity prices; Construct an energy flow model for the entire mining area to achieve coordinated optimization of electricity and fuel.

[0016] As a preferred technical solution of the present invention, the digital twin is constructed through Unity3D, which maps the energy consumption status of the equipment in real time and simulates the impact of different energy-saving strategies on energy consumption. The energy consumption status includes the equipment energy consumption heat map and the energy flow trajectory of the production line.

[0017] As a preferred technical solution of the present invention, in the prediction-control closed-loop feedback step, the deviation rate between the actual energy consumption of the equipment and the predicted value is used as the optimization target, and the parameters of the adaptive prediction model and the global energy-saving strategy are iteratively updated through the gradient descent algorithm.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a sensor cluster to synchronously collect multiple types of data and construct a three-level data grid. It combines noise reduction and missing repair technology to improve data quality, lay the foundation for high-precision prediction, and significantly improve the accuracy of energy consumption prediction.

[0019] 2. This invention uses the hidden Markov model and attention mechanism to accurately identify the dynamic working conditions of equipment and extract key features. Combined with the cross-device transfer learning framework, it can quickly adapt to different types of equipment without large-scale sampling, enhancing the model's generalization ability and deployment efficiency.

[0020] 3. This invention realizes real-time prediction and abnormal warning on the device side through lightweight model deployment, combines the global energy-saving strategy and digital twin visualization on the cloud side, reduces the equipment idling rate and unit energy consumption, forms a "prediction-control-correction" closed loop, and significantly improves the real-time and adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 It is the overall process architecture diagram of the present invention; Figure 2 is a schematic diagram of the edge-cloud collaborative architecture of the present invention; Figure 3 It is a data preprocessing schematic diagram of the present invention; Figure 4 It is a flow chart of dynamic working condition identification and feature extraction of the present invention; Figure 5 is the cross-device transfer learning graph of the present invention; Figure 6 It is a schematic diagram of the prediction-control closed-loop feedback of the present invention; DETAILED DESCRIPTION

[0022] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0023] Example 1 like Figure 1-6 As shown, the present invention provides a method for predicting energy consumption of mining equipment, comprising the following steps: Multi-dimensional data collection: A sensor cluster collects real-time operating data, production process data, and environmental parameters of mining equipment. The operating data includes equipment vibration, temperature, current, rotation speed, and feed rate. The production process data includes blasting parameters, ore grade, and ore size distribution. The environmental parameters include humidity, air pressure, and ore hardness. Data preprocessing: De-noise and repair missing data from the collected raw data. Generative adversarial networks combined with empirical mode decomposition are used to separate valid signals from noise. Missing data is reconstructed based on the correlation between the spatial location of the device and the time series using a spatiotemporal interpolation network. Dynamic operating condition identification and feature extraction: Hidden Markov models are used to identify equipment operating conditions in real time, including excavation, transportation, no-load, full-load, half-load, and standby. Key energy consumption characteristics under different operating conditions are extracted using an attention mechanism. Adaptive prediction model construction: Build a cross-device transfer learning framework to train a universal energy consumption prediction model in the cloud using historical data from multiple mines and multiple equipment models. Then, use a small amount of local data from the target equipment to fine-tune the parameters of the universal model. Edge-cloud collaborative prediction and optimization: A lightweight prediction model is deployed on the device side for real-time energy consumption prediction and abnormal warning. In the cloud, a model predictive control algorithm is used based on data from the entire mine area to generate a global energy-saving strategy. The lightweight model compresses the parameters of the general model through knowledge distillation. Prediction-control closed-loop feedback: Map prediction results to device control parameters, visualize energy consumption status and strategy execution effects through digital twins, analyze model decision-making based on SHAP values, and automatically correct prediction models and control strategies.

[0024] Furthermore, the sensor cluster includes low-cost MEMS sensors and ore hardness testers, which realize the synchronous collection of operation data, production process data and environmental parameters and the construction of a three-level data grid through the Internet of Things. The three-level data grid covers single equipment points, regions and the entire mining area.

[0025] Furthermore, in the data preprocessing step, the generative adversarial network includes a noise generator and a signal discriminator. Through adversarial training, the signal discriminator distinguishes between real and effective signals and pseudo-noise signals generated by the noise generator. Empirical mode decomposition performs multi-scale decomposition on the real and effective signals to extract the intrinsic mode function.

[0026] Furthermore, the spatiotemporal interpolation network constructs the device spatial position association matrix based on the graph neural network, and combines it with the long short-term memory network to process time series data, realizing the reconstruction of missing data through spatiotemporal joint modeling.

[0027] Furthermore, in the dynamic working condition identification and feature extraction steps, the observation sequence of the hidden Markov model is the time series feature vector of the equipment operation data. The transition probability of each working condition is calculated through the forward-backward algorithm. The attention mechanism assigns weights to the energy consumption characteristics under different working conditions to highlight the key influencing factors.

[0028] Furthermore, the cross-device transfer learning framework includes: A general feature extraction layer based on a deep neural network, trained with multi-source device data to learn common energy consumption characteristics across devices; The output layer specific to the target device fine-tunes the parameters of the general feature extraction layer using the local data of the target device to form a prediction model adapted to the specific device.

[0029] Furthermore, in the edge-cloud collaboration step, the parameter scale of the lightweight prediction model is compressed by 60%-80% compared with the original deep learning model, supports localized prediction under offline conditions, and can continue to run for ≥72 hours when the network is disconnected.

[0030] Furthermore, the global energy-saving strategy includes: Real-time planning of the optimal driving path for dump trucks based on ore grade and transportation routes; Adjust the start and stop sequence of high-power equipment, such as air compressors and fans, based on real-time electricity prices; Construct an energy flow model for the entire mining area to achieve coordinated optimization of electricity and fuel.

[0031] Furthermore, the digital twin is built using Unity3D to map the energy consumption status of the equipment in real time and simulate the impact of different energy-saving strategies on energy consumption. The energy consumption status includes the equipment energy consumption heat map and the energy flow trajectory of the production line.

[0032] Furthermore, in the prediction-control closed-loop feedback step, the deviation rate between the actual energy consumption of the equipment and the predicted value is used as the optimization target, and the parameters of the adaptive prediction model and the global energy-saving strategy are iteratively updated through the gradient descent algorithm.

[0033] Specifically, 1. Multi-dimensional data collection steps 1.1 Sensor Deployment and IoT Networking MEMS nine-axis sensors are installed in key parts of mining equipment such as crushers, excavators, and dump trucks. They integrate accelerometers, gyroscopes, and magnetometers to collect vibration and posture data. Current transformers and temperature sensors are deployed to monitor motor current and bearing temperature. Speed ​​encoders and weighing sensors are installed to obtain speed and feed rate.

[0034] Pressure sensors are installed around the blasting holes to collect blasting shock wave parameters; X-ray fluorescence spectrometers are deployed on the ore conveyor belt to detect ore grade in real time, such as the content of elements such as Fe and Cu; and a laser particle size analyzer is installed at the entrance of the crushing chamber to measure the ore size distribution.

[0035] A meteorological station is deployed in the mining area, which includes temperature, humidity and air pressure sensors; an ultrasonic hardness tester is installed in the ore yard to measure the hardness of the ore using the rebound method.

[0036] Afterwards, a hybrid Zigbee+LoRa network was adopted. Device-side sensors were connected to the edge gateway via Zigbee, while environmental and process sensors were transmitted over long distances via LoRa. Finally, data was aggregated to the cloud data center via 5G / wired networks. A three-level data grid was constructed: Equipment level: Real-time operating data of a single device, such as the current waveform of a crusher motor; Regional level: Collaborative data of multiple devices in the same work area, such as the timing coordination of an excavation-transportation equipment group; Mine-level: Mine-wide production plans, such as blasting plans, ore transportation routes, and energy data, such as real-time grid electricity prices and diesel inventory.

[0037] 1.2 Synchronous Collection Using a timestamp alignment algorithm and the device-side edge gateway as the clock source, we synchronize the timestamps of high-frequency data such as vibration and current with low-frequency process data to generate a unified time series data set: ,in: is the operating data of equipment i at time t, including ore grade and fragmentation distribution; Process data, including ore grade and fragmentation distribution; are environmental parameters, including temperature and ore hardness.

[0038] 2. Data Preprocessing Steps 2.1 Noise Robustness Processing Generate adversarial network structure: Noise generator G: Input Gaussian noise z-N(0,1) and generate pseudo noise signal through a 3-layer fully connected network ; Signal discriminator D: input the real signal st or , the CNN output of the three-layer convolution kernel is the discrimination probability P∈[0,1].

[0039] Adversarial training objectives: + Train until D cannot distinguish between real signal and pseudo noise. is the denoised signal.

[0040] For denoised signal Perform empirical mode decomposition to obtain n intrinsic mode functions (IMF1, IMF2, ...IMFn). Retain the first k IMF components with energy ≥ 5% and reconstruct the effective signal: .

[0041] 2.2 Spatiotemporal Interpolation Network Construct the device spatial position adjacency matrix A. If the distance between device i and j is less than 50m, then = 1, otherwise 0. The inter-device spatial weight is calculated through the GCN layer: ; in is the l-layer feature of device j, is the weight matrix.

[0042] Time series data for each device Input long short-term memory network to extract time features .

[0043] Joint spatiotemporal interpolation: The spatial features and time characteristics Splicing, outputting the estimated value of the missing position through the fully connected layer , the loss function is the mean square error; .

[0044] 3. Dynamic working condition identification and feature extraction steps 3.1 Hidden Markov Model Working Condition Identification The equipment working state set S = {excavation, transportation, no load, full load, half load, standby}, a total of 6 states, state transition matrix: ,in .

[0045] The characteristic vector Ot=[current effective value, vibration root mean square value, speed percentage, feeding amount] is extracted from the preprocessed data as the observation sequence of the hidden Markov model.

[0046] The Baum-Welch algorithm is used to estimate the parameters of the hidden Markov model (A, B, π), where B is the observation probability matrix and π is the initial state probability. In real-time recognition, the forward algorithm is used to calculate the probability of each state at time t. , determine the most likely current operating condition .

[0047] 3.2. Attention Mechanism Feature Weighting For current working conditions The corresponding feature sequence , calculate the query vector , key vector , value vector , the weight coefficient is: ; in It is the key vector dimension, highlighting key features, such as the strong correlation between “feed amount” and “motor current” when the crusher is fully loaded.

[0048] 4. Steps to build an adaptive prediction model A multi-layer Transformer encoder is used as a universal model. The multi-device time series data X=[x1,x2,...,xt] is input and the cross-device common energy consumption characteristics are learned through the self-attention mechanism. The pre-training data comes from more than 10 mines and more than 50 models of equipment, with a sample size of ≥ 1 million. The loss function is: .

[0049] Freeze the parameters of the general feature extraction layer and add a device-specific fully connected layer ,enter Output target device energy consumption prediction value Fine-tuning is performed using 2000+ local samples from the target device, and the loss function is: , where λ is the weight decay coefficient to prevent overfitting.

[0050] 5. Edge-Cloud Collaboration Steps 5.1. Lightweight edge model for knowledge distillation, using the teacher-student model architecture: Teacher model: a complex Transformer model pre-trained on the cloud; Student model: lightweight LSTM at the edge.

[0051] Distillation loss function: , where α=0.7, is the soft output of the teacher model, Output of the student model.

[0052] Edge nodes are deployed using ARM Cortex-A72 processors and the Ubuntu IoT system. They store 72 hours of historical data locally and can continue predictions based on the last known operating conditions and local models when the network is disconnected.

[0053] 5.2. Global optimization in the cloud Multi-objective optimization objective function: ; Where: Et is the energy consumption of the equipment at the moment; Ct is the peak-valley electricity price cost, which is dynamically adjusted according to the real-time electricity price of the power grid; Rt is the penalty term for equipment load fluctuation, which prevents frequent start and stop; w1=0.6, w2=0.3, w3=0.1 are weight coefficients.

[0054] The constraints include the upper limit of equipment power, production plan deadline, and the shortest distance constraint of ore transportation routes.

[0055] 6. Prediction-control closed-loop feedback steps 6.1. Digital Twin Visualization 3D Modeling: A 1:1 digital twin of the mining area was built using Unity3D, and the equipment model integrated real-time energy consumption data. For example, the crusher motor temperature is displayed as a heat map. The darker the color, the higher the energy consumption. The energy flow trajectory of the production line is visualized through a particle system. Red particles represent electrical energy and blue particles represent fuel energy.

[0056] Different energy-saving strategies are input into the digital twin environment, such as increasing the crusher discharge opening size by 5mm. The ore crushing process is simulated through a physical engine, the energy consumption change range is predicted, and the optimal strategy is selected.

[0057] 6.2 Model interpretation and parameter iteration: Calculate the contribution of each input feature to energy consumption prediction. For example, if the SHAP value at a certain moment shows "ore hardness + 10HLD → energy consumption + 5KWh", it helps engineers locate the cause of the anomaly.

[0058] Closed-loop optimization algorithm: Defining the Deviation Rate , when δ>15%, the model update is triggered: ; Where θ is the model parameter, η is the learning rate, and the Adam optimizer is used to iterate 50 rounds until δ ≤ 10%.

[0059] Through the above technical solution, after field testing, the accuracy of equipment energy consumption prediction was improved from 80% of the traditional method to 92.5%, the equipment idling rate was reduced by 20%, the unit energy consumption of the crushing link was reduced by 15%, and the response time to abnormal working conditions was shortened from 3 minutes to 15 seconds, verifying the significant improvement of the present invention in accuracy, energy saving effect and real-time performance.

[0060] The comparison between the present invention and traditional statistical models or basic machine learning methods based on single-dimensional data is shown in the following table: In summary, the present invention forms a full-process closed loop of "data acquisition-preprocessing-modeling-collaboration-feedback" through technological innovation, which not only solves the core defects of the existing technology, but also achieves a breakthrough improvement in energy consumption prediction accuracy and system adaptability in practical applications, providing an efficient and reliable technical solution for energy saving and consumption reduction in intelligent mines.

[0061] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and 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 method for predicting energy consumption of mining equipment, characterized in that: The following steps are included: Multi-dimensional data collection: A sensor cluster collects real-time operating data, production process data, and environmental parameters of mining equipment. The operating data includes equipment vibration, temperature, current, rotation speed, and feed rate. The production process data includes blasting parameters, ore grade, and ore size distribution. The environmental parameters include humidity, air pressure, and ore hardness. Data preprocessing: De-noise and repair missing data from the collected raw data. Generative adversarial networks combined with empirical mode decomposition are used to separate valid signals from noise. Missing data is reconstructed based on the correlation between the spatial location of the device and the time series using a spatiotemporal interpolation network. Dynamic operating condition identification and feature extraction: Hidden Markov models are used to identify equipment operating conditions in real time, including excavation, transportation, no-load, full-load, half-load, and standby. Key energy consumption characteristics under different operating conditions are extracted using an attention mechanism. Adaptive prediction model construction: Build a cross-device transfer learning framework to train a universal energy consumption prediction model in the cloud using historical data from multiple mines and multiple equipment models. Then, use a small amount of local data from the target equipment to fine-tune the parameters of the universal model. Edge-cloud collaborative prediction and optimization: A lightweight prediction model is deployed on the device side for real-time energy consumption prediction and abnormal warning. In the cloud, a model predictive control algorithm is used based on data from the entire mine area to generate a global energy-saving strategy. The lightweight model compresses the parameters of the general model through knowledge distillation. Prediction-control closed-loop feedback: Map the prediction results to the device control parameters, and visualize the energy consumption status and strategy execution effect through the digital twin. Analyze the decision basis of the model based on the SHAP value, and automatically correct the prediction model and control strategy.

2. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: The sensor cluster includes low-cost MEMS sensors and ore hardness testers, which use the Internet of Things to achieve the synchronous collection of operating data, production process data and environmental parameters and the construction of a three-level data grid. The three-level data grid covers single equipment points, regions and the entire mining area.

3. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: In the data preprocessing step, the generative adversarial network includes a noise generator and a signal discriminator. Through adversarial training, the signal discriminator distinguishes between real and effective signals and pseudo-noise signals generated by the noise generator. Empirical mode decomposition performs multi-scale decomposition on the real and effective signals to extract intrinsic mode functions.

4. A method for predicting energy consumption of mining equipment according to claim 3, characterized in that: The spatiotemporal interpolation network constructs a device spatial position association matrix based on a graph neural network, and combines it with a long short-term memory network to process time series data, realizing missing data reconstruction through spatiotemporal joint modeling.

5. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: In the dynamic working condition identification and feature extraction step, the observation sequence of the hidden Markov model is the time series feature vector of the equipment operation data. The transition probability of each working condition is calculated through the forward-backward algorithm. The attention mechanism weights the energy consumption characteristics under different working conditions to highlight the key influencing factors.

6. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: The cross-device transfer learning framework includes: A general feature extraction layer based on a deep neural network, trained with multi-source device data to learn common energy consumption characteristics across devices; The output layer that is specific to the target device fine-tunes the parameters of the general feature extraction layer using the local data of the target device to form a prediction model that is adapted to the specific device.

7. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: In the edge-cloud collaboration step, the parameter scale of the lightweight prediction model is compressed by 60%-80% compared with the original deep learning model, supports localized prediction under offline conditions, and can run continuously for ≥72 hours when the network is disconnected.

8. A method for predicting energy consumption of mining equipment according to claim 7, characterized in that: The global energy-saving strategy includes: Real-time planning of the optimal driving path for dump trucks based on ore grade and transportation routes; Adjust the start and stop timing of high-power equipment based on real-time electricity prices; Construct an energy flow model for the entire mining area to achieve coordinated optimization of electricity and fuel.

9. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: The digital twin is built using Unity3D to map the energy consumption status of the equipment in real time and simulate the impact of different energy-saving strategies on energy consumption. The energy consumption status includes the equipment energy consumption heat map and the energy flow trajectory of the production line.

10. A method for predicting energy consumption of mining equipment according to claim 1, characterized in that: In the prediction-control closed-loop feedback step, the deviation rate between the actual energy consumption of the equipment and the predicted value is used as the optimization target, and the parameters of the adaptive prediction model and the global energy-saving strategy are iteratively updated through the gradient descent algorithm.

Images