System for measuring investment intensity of typical low-carbon projects

Abstract

The present application relates to the technical field of low-carbon engineering carbon emission intelligent management, and discloses a typical low-carbon engineering carbon reduction investment intensity calculation system, which collects operation data of low-carbon engineering such as offshore photovoltaic through multi-sensor redundancy, uses blockchain technology to ensure reliable data storage, uses a deep learning model to predict environmental and market factors, dynamically allocates system power based on a multi-objective optimization algorithm, automatically calculates carbon reduction investment intensity in combination with investment cost and carbon emission reduction, and realizes visual presentation and intelligent early warning through digital twin technology. The present application solves the technical problems of inaccurate low-carbon engineering carbon reduction benefit evaluation and difficult quantification of investment returns, and improves the scientificity and accuracy of low-carbon engineering investment decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent carbon emission management technology for low-carbon projects, and more specifically, to a system for calculating the carbon reduction investment intensity of typical low-carbon projects. Background Technology

[0002] Low-carbon projects such as offshore photovoltaic power generation and seawater desalination have been widely adopted in recent years as important technological pathways for reducing carbon emissions. However, the construction of low-carbon projects often requires substantial financial investment, and how to assess the relationship between the carbon reduction benefits and investment costs of these projects has become a crucial issue for policymakers.

[0003] In existing technologies, the assessment of carbon reduction benefits of low-carbon projects mainly relies on simplified static calculation methods, which cannot fully consider dynamic influencing factors such as environmental fluctuations, market price changes, and equipment efficiency degradation, leading to significant discrepancies between the calculated results and actual operating effects. Furthermore, traditional assessment methods generally suffer from technical deficiencies such as low data reliability, poor prediction accuracy, limited optimization strategies, and incomplete assessment dimensions. For example, current power scheduling in combined offshore photovoltaic and seawater desalination systems often relies on fixed rules or manual experience, making it difficult to adapt to the complex and ever-changing marine environment; carbon reduction calculations are usually based on theoretical values ​​or static parameters, failing to reflect actual operating conditions; and investment cost accounting often ignores factors such as equipment performance changes and market fluctuations.

[0004] Therefore, there is an urgent need for a technical solution that can collect operational data in real time, ensure data reliability, accurately predict environmental changes, dynamically optimize power allocation, and scientifically calculate the intensity of carbon reduction investment, in order to solve the technical problems of inaccurate assessment of carbon reduction benefits of low-carbon projects and difficulty in quantifying investment returns. Summary of the Invention

[0005] This invention provides a system for calculating the carbon reduction investment intensity of typical low-carbon projects, solving technical problems such as inaccurate assessment of carbon reduction benefits and difficulty in quantifying benefits in related technologies.

[0006] This invention provides a typical low-carbon engineering carbon reduction investment intensity calculation system, including: The data acquisition and preprocessing module is used to collect raw data from low-carbon engineering sites and perform preprocessing. A blockchain trusted storage module is used to ensure the integrity and traceability of the original data and prevent data from being tampered with. The model prediction module predicts future environmental and market factors based on raw data, providing a basis for decision-making in dynamic power allocation; The dynamic power allocation module is used to dynamically optimize the system power allocation based on model prediction results, thereby maximizing carbon reduction benefits. The carbon emission reduction benefit assessment module is used to calculate the ratio of carbon emission reduction to investment cost of low-carbon projects in real time, and to quantify the carbon reduction investment intensity index. The intelligent early warning and decision support module is used to monitor the system's operating status, identify potential problems, and provide decision suggestions.

[0007] In a preferred embodiment, the data acquisition and preprocessing module includes: Multi-sensor redundant data acquisition, including primary and backup sensor pairs, is used to alternately acquire information on photovoltaic power output, wave height, wave cycle, solar radiation intensity, seawater salinity, park load, and electricity and carbon prices. Multi-link transmission convergence includes three transmission links: submarine cable, satellite, and 4G / 5G, and automatically selects the optimal transmission path through a weighting function; Time synchronization is achieved by employing a dual time synchronization mechanism of GPS and PTP to ensure time consistency across all data collection points. The format is standardized; data from different sources and in different formats is uniformly converted into standard JSON format through a protocol gateway. For missing and outlier handling, a sliding window statistical method combined with machine learning is used to fill in missing data and correct outlier data.

[0008] In a preferred embodiment, the blockchain trusted storage module includes: Data hashing uses the SHA-256 algorithm to calculate the hash value of the time-series dataset and calls a smart contract to upload the hash value to the blockchain network; The consensus nodes adopt a consortium blockchain architecture and select either PBFT or Raft consensus algorithm based on the characteristics of the deployment environment to ensure that all nodes reach a consensus on the record of data hash. The on-chain index uses a B+ tree structure to organize the timestamp index and maintains a mapping table from hash values ​​to transaction IDs, supporting efficient time range queries and hash value queries.

[0009] In a preferred embodiment, the model prediction module includes: Federated learning training uses the FedAvg algorithm to collaboratively train the prediction model on multiple nodes. The prediction model structure adopts a deep learning architecture combining LSTM and attention mechanism. Model compression deployment involves deploying models to edge computing devices through network pruning and parameter quantization. Online forecasting: A forecasting task is triggered every hour, taking in cleaned raw data from the past 168 hours to generate a forecast sequence for the next 24 hours. Interpretable outputs are analyzed using a shape-additive interpretation method to assess the factor contribution of each predicted value. Adaptive fine-tuning is triggered by the mean absolute percentage error (MAPE) of the model.

[0010] In a preferred embodiment, the dynamic power allocation module includes: Initial solution generation: Based on photovoltaic power generation priority, de-emphasis on load matching, electricity price response, carbon price response, and safety margin rules, a feasible initial scheduling scheme is quickly generated. Deep optimization employs an improved particle swarm optimization algorithm or actor-critic reinforcement learning algorithm to generate a better scheduling scheme. The optimization objective function is: ; in, This indicates optimizing the objective function value. Indicates a time step. Indicates carbon cost weighting. Indicates the weight of economic gains. This represents the carbon emission reduction per unit of water produced at time t. This represents the carbon emission reduction at time t. This represents the water production at time t. This represents the economic gain at time t; Digital twin verification involves importing the optimized candidate solutions into a digital twin model for simulation verification. Fault switching automatically switches to a predefined fixed-mode backup strategy in the event of equipment malfunction or communication interruption.

[0011] In a preferred embodiment, the carbon emission reduction benefit assessment module includes: Carbon emission reduction calculation, used to calculate the system's daily cumulative carbon reduction, is calculated using the following formula: ; in, This indicates the cumulative daily carbon reduction. This represents the optimized photovoltaic power at time t. The carbon emission reduction factor representing photovoltaic power generation; This represents the water production at time t. The carbon emission factor representing seawater desalination; Investment cost calculation: Calculate the total investment cost of the system, including equipment purchase cost, installation and commissioning cost, and operation and maintenance cost; The investment intensity index is generated by calculating the carbon reduction investment intensity index based on the daily cumulative carbon reduction and the total investment cost.

[0012] In a preferred embodiment, the formula for calculating the carbon reduction investment intensity by the investment intensity index generation unit is as follows: ; in, This indicates the cumulative daily carbon reduction. Indicates the total investment cost. This indicates the intensity of carbon reduction investment, expressed in tons of carbon dioxide per 10,000 yuan per year.

[0013] In a preferred embodiment, the intelligent early warning and decision support module includes: Threshold monitoring involves regularly scanning system operation data to identify discrepancies in carbon reduction, abnormal investment intensity, and abnormal price fluctuations. Optimization suggestion: Regenerate the scheduling scheme by calling the dynamic power allocation module based on the anomaly type; Notification push notifications will send early warning information and optimization suggestions to relevant personnel via SMS, email, and mobile push channels.

[0014] In a preferred embodiment, the carbon reduction difference calculated by the threshold monitoring unit is: ; in, This indicates differences in carbon reduction. This indicates the actual carbon reduction. This indicates the planned carbon reduction amount.

[0015] In a preferred embodiment, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of the aforementioned typical low-carbon engineering carbon reduction investment intensity calculation system.

[0016] The beneficial effects of this invention are as follows: By employing technologies such as multi-sensor redundant acquisition, multi-link transmission fusion, and blockchain trusted storage, the accuracy and reliability of data have been improved, providing a true and reliable data foundation for the calculation of carbon reduction investment intensity. A deep learning model trained using federated learning enables accurate prediction of environmental and market factors. Combined with a multi-objective optimization algorithm, the system power is dynamically allocated, which significantly improves the system's operating efficiency and carbon reduction benefits. A complete investment intensity accounting method has been constructed, which comprehensively considers multiple factors such as carbon emission reduction and investment costs, and realizes the scientific quantification of the benefits of carbon reduction investment. By using digital twin technology to visualize the system's operational status and combining it with an intelligent early warning mechanism to provide decision support, the system's operability and decision support capabilities are effectively enhanced. The system's scalability and operational efficiency were improved through distributed architecture and heterogeneous computing; data transmission latency and storage pressure were reduced by edge computing technology; long-term prediction accuracy was ensured by AI models based on adaptive fine-tuning; resource allocation efficiency was improved through data-driven dynamic optimization; fault tolerance and reliability were enhanced by intelligent early warning mechanisms; and human-computer interaction experience was improved by immersive visualization technology. Attached Figure Description

[0017] Figure 1 This is a module diagram of the typical low-carbon engineering carbon reduction investment intensity calculation system of the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a typical low-carbon engineering carbon reduction investment intensity calculation system, such as... Figure 1 As shown, it includes: The data acquisition and preprocessing module is used to collect raw data from low-carbon engineering sites and perform preprocessing. In one embodiment of the present invention, the raw data stream mainly covers information such as photovoltaic power output, wave height, wave cycle, solar radiation intensity, seawater salinity, park load, and electricity and carbon prices; the specific data acquisition and preprocessing includes the following steps: S101, multi-sensor redundant acquisition; The system achieves redundant data acquisition through sensor groups; the sensor groups include primary and backup sensor pairs. Taking the photovoltaic output sensor as an example, the system deploys two independent photovoltaic output monitoring devices at the same time, which are referred to as the primary sensor and the backup sensor, respectively; these sensors alternately collect data according to a preset sampling frequency (e.g., once every 5 minutes); When a primary sensor malfunction is detected (e.g., three consecutive abnormal sampling values ​​or no response), the system automatically switches to the backup sensor and issues a primary sensor maintenance reminder. This redundant acquisition mechanism ensures the continuity of data acquisition in the event of a single point of failure. S102, multi-link transmission convergence; The collected raw data is transmitted via a data transmission network, which includes submarine cables, satellite, and 4G / 5G transmission links. The system monitors the transmission quality indicators (including latency, packet loss rate, and bandwidth) of each link in real time and automatically selects the optimal transmission path based on a preset weighting function. The weighting function can be expressed as: ; in, This represents the overall weight of the i-th link. This represents the normalized delay of the i-th link. This represents the normalized packet loss rate of the i-th link. This represents the normalized bandwidth of the i-th link. , and Let these represent the weighting coefficients for latency, packet loss, and bandwidth, respectively, and satisfy the following conditions: ;choose The shortest link is selected as the current optimal transmission path (i.e., the path with the highest transmission efficiency).

[0020] S103, time synchronization; A dual time synchronization mechanism using GPS and PTP (Precise Time Protocol) is employed to ensure time consistency across all data collection points. Each data collection point is equipped with a GPS receiver as the primary time source, while the PTP protocol is run over the network as a backup time synchronization mechanism. The difference between the two time sources is compared periodically, and when the difference exceeds a preset threshold (e.g., 5ms), an alarm is triggered and manual intervention is required. Through this dual time synchronization mechanism, the system ensures that the time synchronization error is controlled within 10ms. S104, standardized format; The protocol gateway converts data from different sources and in different formats into a unified standard JSON format. The protocol gateway supports mainstream industrial protocols such as Modbus, OPCUA, and MQTT, and maps data points from various protocols to a unified JSON field according to a pre-configured mapping table. Taking the Modbus protocol as an example, the gateway maps the Modbus register address to a specific JSON field name and performs appropriate conversion according to the data type; the unified JSON data contains necessary metadata (such as timestamp, source identifier, quality code, etc.) and actual measurement values.

[0021] S105, Missing and Exception Handling; The system performs missing data detection and anomaly identification on the standardized data; it also uses a sliding window statistical method combined with machine learning to check data quality. Specifically, the system maintains a historical data window of length n (e.g., n=24, representing 24 hours), and calculates the statistical characteristics (mean, standard deviation, rate of change, etc.) of the data within the window; for newly collected data points, the system calculates their outlier scores using a pre-trained isolated forest algorithm. ; in, This represents the anomaly score of sample x. This represents the average path length of sample x. Indicates the expected path length. is the standardization factor, and n is the window length. When When the value exceeds a preset threshold (e.g., 0.7), it is identified as an anomaly.

[0022] For detected missing data, the system employs different imputation strategies based on the data type: For slowly varying data (such as seawater salinity), linear interpolation is used to fill in the gaps. For periodic data (such as photovoltaic power output), the average value of the same historical period is used to fill the gaps. For random data (such as electricity prices), fill in the blanks with the previous valid value; For any abnormal data detected, the system first attempts to switch to backup sensor data; if the backup data is also abnormal, an online calibration procedure is initiated to correct the abnormal values ​​based on a historical data model. The specific correction steps are as follows: A historical data model is constructed. Based on the effective raw data of the same period in the past 30 days, the system establishes an ARIMA time series prediction model. Anomaly assessment uses the standardized root mean square error (NRMSE) method to calculate the deviation between outliers and model predictions, and determines correction weights based on the magnitude of the deviation. Weighted correction involves taking a weighted average of outliers and historical data model predictions according to calculated weights to generate corrected data values. Confidence markers are added to the corrected data for reference by subsequent modules. The process of building a historical data model includes: Data segmentation categorizes low-carbon engineering field data such as photovoltaic output, ocean wave height / cycle, and solar radiation intensity according to characteristics such as time period and weekday / non-working day. Different segmentation strategies are adopted for slowly varying, periodic, and random data. Feature extraction: Extracting trend, periodic and autocorrelation features from various sensor data, with particular attention to features such as the correlation between photovoltaic output and solar radiation intensity, and seasonal changes in seawater salinity; Parameter optimization involves determining the optimal parameters (p, d, q) of the ARIMA model (autoregressive integral moving average model used for analyzing and forecasting time series data) through grid search (an exhaustive parameter optimization technique that systematically tries all possible parameter combinations in a predefined parameter space and evaluates the model performance of each set of parameters). Where p represents the number of autoregressive terms (the degree of influence of past values ​​on the current value), d represents the difference order (the number of differences required to transform a non-stationary time series into a stationary series), and q represents the number of moving average terms (the degree of influence of past prediction errors on the current value). Different parameter settings are used for different types of sensor data. For example, photovoltaic power output data may require higher-order differential processing (d value) to eliminate seasonal fluctuations and trend changes, ocean wave height data may require more autoregressive terms (p value) to capture its periodic change patterns, and electricity price data may require more moving average terms (q value) to smooth random fluctuations. Model validation is performed using k-fold cross-validation (k=5) with data from the same period over the past 30 days. The dataset is divided into 5 subsets. The model is trained on 4 subsets and tested on the remaining 1 subset each time. After 5 iterations, the average performance index is taken to evaluate the prediction accuracy of the model in different time windows and to ensure the stability and reliability of the model in the process of correcting abnormal data. The data acquisition and preprocessing module outputs a high-quality time-series dataset that has undergone redundant acquisition from multiple sensors, multi-link transmission fusion, time synchronization, format unification, and handling of missing and anomalies. This dataset includes the following key fields: Data collection timestamps are accurate to the millisecond level and synchronized using both GPS and PTP time synchronization mechanisms; The output power of the photovoltaic system has undergone anomaly detection and correction. Wave height is used to assess the potential for wave energy generation. The wave cycle, together with the wave height, determines the wave energy conversion efficiency. Solar radiation intensity directly affects photovoltaic power generation efficiency; Seawater salinity affects the energy consumption and equipment lifespan of seawater desalination. The park's electricity load reflects the energy demand-side situation. Electricity prices reflect the state of the energy market; Carbon price is used for carbon emission cost accounting; Each data record is accompanied by a quality code indicating whether the data is a raw collected value, an interpolated filled value, or a model-corrected value, along with the corresponding confidence score.

[0023] A blockchain trusted storage module is used to ensure the integrity and traceability of the original data and prevent data from being tampered with. In one embodiment of the present invention, the processing procedure of the blockchain trusted storage module includes the following steps: S201, Data Hashing and Packaging; The data hashing unit performs hashing processing on the received time-series dataset; The system aggregates the hourly time-series dataset into a single data block every hour and calculates the hash value of that block. The hash algorithm used is SHA-256 to ensure data integrity and immutability. The calculation formula is as follows: in, This represents the timestamp of the i-th data point. This represents the data content of the i-th data point, where n represents the number of data points. This indicates a string concatenation operation, where n represents the total number of data points.

[0024] After calculating the hash value, the data hash unit calls the smart contract uploadData(hash, timestamp) to upload the hash value and the corresponding timestamp to the blockchain network. uploadData is a smart contract function used to upload data to the blockchain network for storage and verification. The basic logic of smart contracts includes permission verification, data format verification, and storage operations, ensuring that only authorized nodes can upload data and that the data format conforms to predefined specifications; S202, consortium blockchain consensus; Consensus nodes participate in the consensus process of the blockchain network; the system adopts a consortium blockchain architecture, including multiple distributed consensus nodes; depending on the characteristics of the deployment environment, the system can choose PBFT (Practical Byzantine Fault Tolerance) or Raft consensus algorithm; For scenarios requiring high security, the system adopts the PBFT algorithm, which can still guarantee system consistency even if up to one-third of the nodes experience Byzantine failure (i.e., nodes may send incorrect information or engage in malicious behavior). The PBFT consensus process includes four phases: request, pre-preparation, preparation, and confirmation. The three-phase commit protocol ensures that all honest nodes reach a consensus. Optionally, for scenarios where performance is more critical, the system employs the Raft algorithm, which operates under the assumption that nodes are honest but may fail, and has high throughput; Raft achieves distributed consensus through leader election, log replication, and security rules. Through these consensus mechanisms, the system ensures that all nodes agree on the record of data hashes, preventing single nodes from tampering with the data; S203, on-chain index; On-chain indexes create and maintain data indexes on the blockchain; whenever a new data hash is confirmed and added to the chain, the system records the block number, transaction ID, data hash value, and corresponding timestamp of the transaction, forming an index record {block number, transaction ID, data hash, timestamp}; these index records are stored in a dedicated index data structure, supporting efficient time range queries and hash value queries; To improve query efficiency, the system uses a B+ tree structure to organize the timestamp index. The B+ tree structure is a balanced binary tree with high search performance, which is particularly suitable for range queries and time series data. It also maintains a mapping table from hash values ​​to transaction IDs. In this way, the system can quickly locate data records in a specific time period or verify whether a specific data hash has been put on the chain and its on-chain position. The blockchain trusted storage module outputs a verified time-series dataset; the on-chain verification record corresponding to the time-series dataset can be queried and verified for data integrity by querying the data function queryData(timestamp) → data hash value.

[0025] The model prediction module predicts future environmental and market factors based on raw data, providing a basis for decision-making in dynamic power allocation; In one embodiment of the present invention, the processing procedure of the model prediction module includes the following steps: S301, Federal Learning and Training; The model training unit uses a federated learning framework to train the prediction model; Federated learning allows multiple nodes to collaboratively train a model without sharing the original data, thus protecting data privacy. Each edge computing node in the system acts as a participant in federated learning, training a local model based on local data, and then uploading only the model parameters (rather than the original data) to the central server for aggregation. The initial input to the model is multidimensional time series data, including features such as photovoltaic power generation, wave height, wave period, solar radiation intensity, seawater salinity, load demand, electricity price, and carbon price. This data is standardized and then input into the network in chronological order. The model structure adopts a deep learning architecture that combines LSTM (Long Short-Term Memory Network) with a recollection mechanism; the LSTM layer is responsible for capturing long-term dependencies in time series data, and its core consists of an input gate, a forget gate, and an output gate; The specific working principle is as follows: The forget gate determines which information needs to be discarded from the previous state; it concatenates the hidden state from the previous time step with the current input, processes it through the sigmoid function, and outputs a value between 0 and 1 to control the degree to which the cell state from the previous time step is preserved. The input gate determines which information needs to be updated; it also uses the previous hidden state and the current input to generate a control value through the sigmoid function, while creating a new candidate cell state (processed through the tanh function). The system updates the cell state by multiplying the previous state by the output of the forget gate (discarding some information) and adding the new candidate state controlled by the input gate (adding new information). The output gate control determines what information to output based on the updated cell state; it uses the sigmoid function to process the combination of the previous hidden state and the current input, and then multiplies it with the cell state processed by the tanh function to generate the hidden state output at the current time step. The attention mechanism is used to highlight important time steps, and its workflow is as follows: The system calculates the energy score for each time step by transforming the current hidden state and the previous decoder state through a weight matrix, applying the tanh function, and then multiplying it with the learnable parameter vector. These energy scores are converted into attention weights using the softmax function, ensuring that the sum of all weights is 1; The system uses these attention weights to sum all hidden states in a weighted manner, generating a context vector for subsequent prediction; The training process of federated learning uses the FedAvg algorithm, which achieves distributed model training by training the model on multiple local nodes and aggregating parameter updates. The system first initializes the global model parameters by the central server, and then distributes the parameters to each participating node. Each node trains the model using local data and calculates the parameter updates, and then sends the updates back to the central server. The central server weights and averages all parameter updates according to the amount of data in each node to generate a new global model. This process is repeated until the model converges or reaches the predetermined number of training rounds. S302, Model Compression Deployment; The trained model is compressed and then deployed to edge computing devices; model compression employs two main techniques: network pruning and parameter quantization. Network pruning reduces model size by removing connections or neurons that contribute less to prediction results. The system calculates the importance score of each connection, removes connections with scores below a threshold, and then fine-tunes the pruned network to restore model performance. Parameter quantization converts 32-bit floating-point parameters into 8-bit fixed-point representations, reducing model storage space and computational complexity. The system first determines the maximum absolute value of each layer's parameters, calculates the scaling factor, and then quantizes the parameters into 8-bit integers. During inference, the approximate parameter values ​​are recovered through dequantization. Through the above compression techniques, the system reduces the model size by about 3 / 4, significantly reduces inference time, and meets the resource constraints of edge computing devices; S303, online prediction; The compressed deployment model is periodically invoked for prediction; the system triggers a prediction task once per hour, inputting the cleaned raw data of the past 168 hours (7 days) to generate a prediction sequence for the next 24 hours.

[0026] The prediction process standardizes the input data, mapping each feature value to the interval [-1, 1]. The system organizes the standardized input data using a sliding window method, with a window size of 24 (representing 1 day) and a step size of 1. For each window, the system outputs the predicted value for the next 24 hours through a forward propagation calculation model. Finally, the system denormalizes the model output to restore it to the original data scale. S304, Interpretable Output; The system uses the Shape Additive Explanations (SHAP) method to analyze the factor contribution of each predicted value. Shape Additive Explanations are based on the Shapley value concept in game theory, calculating the marginal contribution of each feature to the model's prediction. The system generates a SHAP interpretation for each predicted value, including a ranking of the contributions of each factor (such as solar radiation intensity and wave height) and a visual chart to help users understand the model's decision-making process. Since the calculation of shape additive interpretation values ​​involves complex combinatorial mathematics, the system adopts an approximation method using a deep shape additive interpretation algorithm. The specific steps are as follows: The system treats deep networks as multi-layer composite functions, calculating them sequentially from the input layer to the output layer; For a given predicted value System computational characteristics Additive interpretation of shape: ; in, The shape-additive interpretation value of feature i; Represents the set of all features. Indicates that it does not contain features A subset of; To represent factorial; This represents the predicted value of a subset containing feature i; This represents the predicted value of the subset that does not contain feature i; In one embodiment of the present invention, since directly calculating the above formula is too computationally complex, the system adopts a recursive method to propagate the contribution from the output layer back layer by layer. For each layer, the system calculates the contribution of each neuron in that layer to the next layer and assigns the contribution to the neurons in the previous layer. Through multiple recursive calculations, the system finally obtains the contribution of each feature in the input layer to the prediction result, i.e., the shape additive interpretation value. The system sorts each input feature (such as solar radiation intensity, wave height, wave period, seawater salinity, etc.) according to its contribution and generates visual representations such as horizontal bar charts.

[0027] This recursive computation method reduces computational complexity, enabling the system to generate interpretable results in real time on edge computing devices. This helps users understand the basis of the model's predictions under different environmental conditions, improving the system's transparency and credibility.

[0028] S305, adaptive fine-tuning; Monitor the model's predictive performance and update the model when necessary (when the mean absolute percentage error between the predicted and observed values ​​exceeds a preset threshold, such as 8%); the system periodically (e.g., daily) compares the predicted and observed values ​​and calculates the mean absolute percentage error (MAPE). ; in, Indicates the mean absolute percentage error. This represents the actual value of the i-th sample. This represents the predicted value of the i-th sample. Indicates the number of samples.

[0029] When MAPE exceeds a preset threshold (e.g., 8%), the system triggers model fine-tuning; The fine-tuning process uses the raw data from the last 24 hours to perform mini-batch gradient descent and update the model parameters. Fine-tuning involves only a small number of iterations (e.g., 5-10 rounds) to balance update speed and computational resource consumption; After fine-tuning, the system re-evaluates the model performance; if the performance improves, the updated model is retained; otherwise, it reverts to the original model and issues an alert, indicating that manual intervention or complete retraining may be necessary. The model prediction module outputs 24-hour prediction data, including prediction data for the next 24 hours (wave height time series, irradiance time series, salinity time series, electricity price time series, carbon price time series) and corresponding explanatory information.

[0030] The dynamic power allocation module is used to dynamically optimize the system power allocation based on model prediction results, thereby maximizing carbon reduction benefits. In one embodiment of the present invention, the processing procedure of the dynamic power allocation module includes the following steps: S401: Initial solution generation; The system quickly generates feasible initial scheduling schemes based on simple rules. It uses a rule engine to process input data and generates initial solutions based on the following heuristic rules: The priority rule for photovoltaic power generation is that during periods of high solar irradiance, photovoltaic power generation is allocated to the maximum value first. De-emphasize load matching rules and strive to match power supply with water demand to ensure water supply. Electricity price response rules stipulate that during peak electricity price periods, descaling power should be appropriately reduced to decrease electricity costs. The carbon price response rule is to increase photovoltaic utilization and reduce carbon emissions when carbon prices are high. Safety margin rule: Retain approximately 15% power margin at each time period to cope with emergencies; Based on the above rules, the system generates an initial scheduling plan for 24 hours, including hourly photovoltaic power allocation and seawater desalination power allocation; the initial solution generation process takes less than 0.5 seconds to ensure that a feasible solution can be generated quickly even in extreme cases. S402, deeply optimized; Using the initial solution as a seed, an iterative optimization algorithm is used to generate a better scheduling scheme. The system provides two optimization algorithms: an improved particle swarm optimization algorithm and an actor-critic reinforcement learning algorithm, which can be selected according to the specific scenario. For the improved particle swarm optimization algorithm, the system represents the scheduling scheme as a 48-dimensional vector (24 hours × 2 control variables), with particle position representing the power allocation scheme and particle velocity representing the update direction of the scheme. The algorithm employs adaptive inertia weights and hybrid acceleration coefficients to enhance global search capabilities. The optimization objective function is: ; in, This indicates optimizing the objective function value. Indicates a time step (hours). Indicates carbon cost weighting. Indicates the weight of economic gains. This represents the carbon emission reduction per unit of water produced at time t. This represents the carbon emission reduction at time t. This represents the water production at time t. This represents the economic gain at time t; Particle swarm optimization (PSO) searches for the optimal solution by iteratively updating particle positions and velocities. In each iteration, the system evaluates the fitness (objective function value) of each particle, updates the individual optimal position and the global optimal position, and then adjusts the particle velocity and position according to the update formula. The system also checks the constraints and corrects particles that violate the constraints; the iterative process continues until the maximum number of iterations is reached or the convergence condition is met, and finally outputs the global optimal position as the optimized scheduling scheme. Optionally, for the actor-critic reinforcement learning algorithm, the system models the scheduling problem as a Markov decision process; the state space contains information such as the current time, environmental predictions, and system parameters; the action space is the allocation value of photovoltaic power and de-flashing power; and the reward function is the optimization objective function. The system uses two neural networks: the actor network is responsible for selecting actions based on the state, and the critic network evaluates the value of the state-action pair; through temporal difference learning and policy gradient methods, the system continuously optimizes the scheduling strategy and finally generates a high-quality scheduling scheme. In one embodiment of the present invention, regardless of which optimization algorithm is selected, the system will perform constraint checks on the generated scheduling scheme to ensure that the photovoltaic power and desalination power do not exceed the maximum capacity and that the water production meets the basic requirements. For example, in a typical summer sunny day's scheduling scheme, the system allocates photovoltaic power to near its maximum capacity (approximately 90%) from 9:00 AM to 3:00 PM, making full use of solar resources. Simultaneously, based on the water demand curve, during peak water usage periods (7:00 AM to 9:00 AM and 6:00 PM to 8:00 PM), the desalination power is increased to over 80% to ensure sufficient water supply. During off-peak nighttime periods (11:00 PM to 5:00 AM the next day), the desalination power is reduced to approximately 30% of the baseline operating level, while reserving approximately 15% of the system capacity as an emergency backup. This dynamic scheduling scheme satisfies basic constraints while achieving efficient resource utilization and minimizing carbon emissions.

[0031] S403, Digital Twin Verification; The optimized candidate solutions are imported into a digital twin model for simulation verification. A digital twin model is a high-precision software replica of the physical system, containing detailed equipment characteristics, system dynamics, and environmental influencing factors.

[0032] The system first applies the candidate scheduling schemes to the digital twin model to simulate one hour of system operation; During the simulation, the system records the actual response curves of key parameters, including the actual output power of the photovoltaic system, the actual power consumption of the desalination system, the water production, and the system losses. Next, the system calculates the deviation between the simulation results and the expectations: ; ; ; in, This represents the relative deviation of photovoltaic power at time t. This represents the simulated photovoltaic power at time t. This represents the planned photovoltaic power at time t; This represents the relative deviation of the fade-out power at time t. This represents the simulated fade-out power at time t. This represents the planned fade-out power at time t; This indicates the relative deviation of water production at time t. This represents the simulated water production at time t. This represents the planned water production at time t.

[0033] S404, fault switching; The system is equipped with fault detection and emergency response mechanisms to ensure that basic operation can be maintained in the event of equipment malfunction or communication interruption; It monitors device status signals and communication link quality in real time, and automatically switches to a predefined "fixed mode" backup strategy when an anomaly is detected (such as device alarm, sensor failure, communication timeout, etc.). The fixed-mode strategy is pre-set based on historical operating data and expert experience, including contingency plans for several typical scenarios, such as "maximum safety mode" (all equipment operates with safe parameters), "load priority mode" (prioritizing water demand), and "power generation priority mode" (prioritizing maximizing photovoltaic utilization). The system selects an appropriate backup strategy based on the type and severity of the fault to ensure safe and stable system operation. Once the system returns to normal, it will automatically switch back to optimized scheduling mode and record the fault events and handling process for subsequent analysis and improvement. The dynamic power allocation module outputs an optimized 24-hour scheduling table, which is an optimized and verified 24-hour scheduling table containing information such as (optimized photovoltaic power time series, optimized desalination power time series, water production time series, and reserved capacity), and sends it to the control system as an execution command. The carbon emission reduction benefit assessment module is used to calculate the carbon emission reduction benefit indicators of low-carbon projects, providing a quantitative basis for investment decisions; In one embodiment of the present invention, the processing procedure of the carbon emission reduction benefit assessment module includes the following steps: S501, Daily Cumulative Carbon Reduction Calculation; The carbon emission reduction calculation unit calculates the system's daily cumulative carbon reduction based on the optimized scheduling scheme; the carbon reduction calculation is based on two parts: The carbon emission reduction brought about by photovoltaic power generation replacing traditional energy sources and the carbon emissions during seawater desalination; the calculation formula is: ; in, This indicates the cumulative daily carbon reduction. This represents the optimized photovoltaic power at time t. The carbon emission reduction factor represents the amount of carbon emissions reduced per kilowatt-hour of photovoltaic power generation compared to traditional energy sources. This represents the water production at time t. The carbon emission factor represents the carbon emissions generated per cubic meter of desalinated seawater. The carbon factor is determined using the life cycle assessment method, which comprehensively considers carbon emissions throughout the entire process of equipment manufacturing, transportation, installation, operation and scrapping; the system supports dynamic adjustment of the carbon factor value based on factors such as equipment batch, operating status and environmental conditions, thereby improving calculation accuracy; S502, Calculation of total investment cost; The total investment cost of the computing system; the investment cost includes equipment purchase cost, installation and commissioning cost, operation and maintenance cost, etc., and the calculation formula is: ; in, Indicates the total investment cost. Indicates the maximum capacity of the photovoltaic system. This indicates the investment cost per unit of photovoltaic capacity; This indicates the maximum capacity of the de-diminishing system. This indicates the investment cost per unit of desalination capacity; represents the regional adjustment coefficient, reflecting cost differences in different regions; m represents the operation and maintenance coefficient, which represents the annual operation and maintenance cost per unit capacity.

[0034] The system supports adjusting various parameters according to the actual situation of the project, taking into account the impact of factors such as equipment depreciation, technological progress, and subsidy policies, to ensure the accuracy of investment cost calculation; at the same time, the system maintains a dynamically updated cost database, recording the latest market prices of equipment from different suppliers and of different specifications, providing a reference for investment cost estimation. S503, the investment intensity index is generated; The carbon reduction investment intensity index is calculated based on the daily cumulative carbon reduction and the total investment cost; the formula for calculating investment intensity is: ; in, This represents the cumulative daily carbon reduction, and 365 represents the number of days in a year. Indicates the total investment cost. This indicates the carbon reduction investment intensity, expressed in tons of carbon dioxide per 10,000 yuan per year. This indicator represents the amount of carbon emissions that can be achieved per 10,000 yuan of investment per year; the higher the value, the better the emission reduction benefits of the investment. The system supports the calculation of investment intensity indicators based on different time scales, including daily, weekly, monthly, and annual indicators, reflecting changes in investment benefits over different time periods. At the same time, the system provides auxiliary financial indicators such as investment payback period and internal rate of return to help decision-makers comprehensively assess the economic and environmental benefits of projects. The carbon emission reduction benefit assessment module outputs a daily report, which includes information such as (date, daily carbon reduction, total investment, and investment intensity).

[0035] The intelligent early warning and decision support module is used to monitor the system's operating status, identify potential problems, and provide decision suggestions. In one embodiment of the present invention, the processing procedure of the intelligent early warning and decision support module includes the following steps: S601, Threshold monitoring; Regularly scan system operation data to identify anomalies. The system should be scanned every 15 minutes, checking the following key metrics: Carbon reduction variance: Calculating the difference between actual carbon reduction and planned values. ,when Timely triggering of warnings; The calculation formula is: ; in, This indicates differences in carbon reduction. This indicates the actual carbon reduction. Indicates the planned carbon reduction amount; If the investment intensity is abnormal, check whether the real-time calculated investment intensity ρ exceeds the preset upper limit. ,when Timely triggering of warnings; Price fluctuations: Examine the fluctuations in market prices (electricity prices, carbon prices), and identify when price changes exceed the standard deviation. The system triggers an alert based on historical price data. ; The specific method is to first calculate the average of the price data over the past 30 days, then calculate the squared difference between each price point and the average, sum these squared differences, divide by the sample size, and finally take the square root to obtain the standard deviation. The system uses price data from the past 30 days to calculate the standard deviation by default and updates it daily. When the absolute value of the difference between the real-time price and the price at the same time the previous day exceeds twice the standard deviation (this multiple is a configurable sensitivity coefficient, with a default value of 2), the system will trigger a price fluctuation warning. The system uses a sliding time window for statistical analysis to reduce the impact of short-term fluctuations; at the same time, the system supports dynamic threshold adjustment, automatically adjusting the warning threshold according to factors such as season and weather to improve the accuracy of warnings. S602: Optimization suggestions; Upon detecting an anomaly, the dynamic power allocation module is invoked to regenerate the scheduling scheme; the system employs different optimization strategies based on the anomaly type: When carbon reduction discrepancies are abnormal, and the actual carbon reduction is significantly lower than the planned value, the system increases the α weight (carbon cost weight) and decreases the β weight (economic benefit weight), re-optimizes the scheduling scheme, and improves carbon reduction efficiency. When investment intensity is abnormal and exceeds the preset upper limit, the system adjusts and optimizes the objective function, increases investment return constraints, and generates a more economical and efficient scheduling scheme. When price fluctuations are abnormal and market prices change significantly, the system re-predicts future trends based on the latest price data, adjusts and optimizes parameters, and adapts to market changes. The optimization suggestions generated by the system include new scheduling schemes, as well as the specific reasons for the adjustments and the expected effects. The suggestions include both short-term countermeasures (such as power adjustments for the remaining time of the day) and medium- to long-term optimization strategies (such as equipment maintenance recommendations and system upgrade directions). S703, push notifications; The system pushes early warning information and optimization suggestions to relevant personnel through multiple channels; it supports multiple notification methods such as SMS, email, and mobile push, and selects the appropriate push method according to the warning level and user role. The push notification includes information such as {warning type, warning time, original scheduling plan, and optimized scheduling plan}. The warning types include differences in carbon reduction, abnormal investment intensity, or price fluctuations. The system supports a tiered push strategy, determining the recipients and methods of push notifications based on the urgency and scope of the alert. For example, general reminders are pushed to maintenance personnel via the app; important alerts are simultaneously pushed to department heads via SMS and email; and emergency alerts are pushed to all relevant personnel through all channels and trigger audible and visual alarms.

[0036] The system also implements an intelligent anti-fatigue mechanism to avoid repeatedly pushing similar warnings in a short period of time, thus reducing the burden on users; When similar warnings are detected to be triggered multiple times in a short period of time, the system will merge these warnings into a single comprehensive message and appropriately reduce the frequency of push notifications. The intelligent early warning and decision support module outputs an early warning log, which includes information such as early warning timestamp, early warning type, original investment intensity, and new 24-hour scheduling plan.

[0037] In one embodiment of the present invention, an application example of a typical low-carbon engineering carbon reduction investment intensity measurement system is provided: To verify the technical effect of the present invention, the proposed carbon reduction investment intensity calculation system was applied to a joint offshore photovoltaic and seawater desalination project. The project includes a 5MW offshore photovoltaic power generation system and a 10,000-ton-per-day seawater desalination system, which is deployed in the waters near an island in the East China Sea. Application testing was conducted in a real-world production environment over a three-month period, covering various weather conditions and load scenarios. During the testing, the system collected real-time data on photovoltaic output, ocean wave parameters, and water production, while also recording fluctuations in electricity and carbon prices. The test environment was equipped with a complete sensor network, including photovoltaic power station energy meters, wave observation buoys, irradiance meters, seawater conductivity meters, load monitoring devices, and electricity / carbon price data interfaces; the data acquisition frequency was set to once every 5 minutes, and more than 100,000 raw data records were collected during the system operation. The system first cleans and preprocesses the collected data, identifying and correcting approximately 3% of outlier data points. The processed data is stored through a blockchain system to ensure its credibility. The model prediction module trains an LSTM+Attention model based on the raw data to predict environmental and market factors for the next 24 hours, with prediction accuracy remaining between 92% and 95% under different weather conditions. Table 1 shows the prediction accuracy of the AI ​​prediction module under different weather conditions. Table 1: Statistics on AI Prediction Accuracy under Different Weather Conditions The dynamic power allocation module uses an improved PSO algorithm to optimize system operating parameters, maximizing carbon reduction benefits while meeting water supply demands. The system-generated scheduling scheme is executed after digital twin verification, with actual operational deviations controlled within ±3%. By comparing traditional fixed scheduling schemes with the dynamically optimized scheduling scheme of this system, key indicator data such as photovoltaic utilization rate and system carbon reduction were collected. The results show that the proposed system scheme has significant advantages.

[0038] Table 2 shows a performance comparison between traditional fixed scheduling and dynamic optimized scheduling. Table 2: Performance Comparison of Traditional Fixed Scheduling and Dynamic Optimized Scheduling The investment intensity index calculated by the carbon emission reduction benefit assessment module is 1.65 tCO2 / 10,000 yuan / year, which clearly quantifies the relationship between the environmental and economic benefits of the project. The system also records the monthly change data of investment intensity during the test period, reflecting the impact of seasonal changes on carbon reduction benefits. Table 3 shows the monthly trend of carbon reduction investment intensity; Table 3: Monthly Trend of Carbon Reduction Investment Intensity Table 4 shows the relationship between solar irradiance and photovoltaic power generation. The analysis results show that the dynamic power allocation strategy of this system can make more effective use of variable solar irradiance resources. Table 4: Analysis of the Relationship between Solar Radiation Intensity and Photovoltaic Power Generation Table 5 shows the relationship between energy consumption and water production of the desalination system, and also shows that the optimized scheduling of this system can reduce energy consumption while ensuring water supply demand. Table 5: Energy Consumption and Water Production of Desalination System Test results show that, compared with the traditional fixed scheduling strategy, the dynamically optimized scheduling scheme of this system significantly improves the photovoltaic utilization rate and the overall system efficiency; under typical sea conditions, the carbon reduction of the system is significantly improved compared with the traditional scheme; at the same time, the water production cost is effectively reduced and the system operation stability is significantly improved; the carbon reduction investment intensity index clearly quantifies the relationship between the environmental and economic benefits of the project, providing a reliable basis for subsequent investment decisions.

[0039] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A typical low-carbon project carbon reduction investment intensity calculation system, characterized in that, include: The data acquisition and preprocessing module is used to collect raw data from low-carbon engineering sites and perform preprocessing. A blockchain trusted storage module is used to ensure the integrity and traceability of the original data and prevent data from being tampered with. The model prediction module predicts future environmental and market factors based on raw data, providing a basis for decision-making in dynamic power allocation; The dynamic power allocation module is used to dynamically optimize system power allocation based on model prediction results to maximize carbon reduction benefits; it includes: Initial solution generation: Based on photovoltaic power generation priority, de-emphasis on load matching, electricity price response, carbon price response, and safety margin rules, a feasible initial scheduling scheme is quickly generated. Deep optimization employs an improved particle swarm optimization algorithm or actor-critic reinforcement learning algorithm to generate a better scheduling scheme. The optimization objective function is: ； in, This indicates optimizing the objective function value. Indicates a time step. Indicates carbon cost weighting. Indicates the weight of economic gains. This represents the carbon emission reduction per unit of water produced at time t. This represents the carbon emission reduction at time t. This represents the water production at time t. This represents the economic gain at time t; Digital twin verification involves importing the optimized candidate solutions into a digital twin model for simulation verification. Fault switching automatically switches to a predefined fixed-mode backup strategy in the event of equipment failure or communication interruption. The carbon emission reduction benefit assessment module is used to calculate the ratio of carbon emission reduction to investment cost of low-carbon projects in real time, and to quantify the carbon reduction investment intensity index. The intelligent early warning and decision support module is used to monitor the system's operating status, identify potential problems, and provide decision suggestions.

2. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 1, characterized in that, The data acquisition and preprocessing module includes: Multi-sensor redundant data acquisition, including primary and backup sensor pairs, is used to alternately acquire information on photovoltaic power output, wave height, wave cycle, solar radiation intensity, seawater salinity, park load, and electricity and carbon prices. Multi-link transmission convergence includes three transmission links: submarine cable, satellite, and 4G / 5G, and automatically selects the optimal transmission path through a weighting function; Time synchronization is achieved by employing a dual time synchronization mechanism of GPS and PTP to ensure time consistency across all data collection points. The format is standardized; data from different sources and in different formats is uniformly converted into standard JSON format through a protocol gateway. For missing and outlier handling, a sliding window statistical method combined with machine learning is used to fill in missing data and correct outlier data.

3. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 1, characterized in that, The blockchain trusted storage module includes: Data hashing uses the SHA-256 algorithm to calculate the hash value of the time-series dataset and calls a smart contract to upload the hash value to the blockchain network; The consensus nodes adopt a consortium blockchain architecture and select either PBFT or Raft consensus algorithm based on the characteristics of the deployment environment to ensure that all nodes reach a consensus on the record of data hash. The on-chain index uses a B+ tree structure to organize the timestamp index and maintains a mapping table from hash values ​​to transaction IDs, supporting efficient time range queries and hash value queries.

4. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 1, characterized in that, The model prediction module includes: Federated learning training uses the FedAvg algorithm to collaboratively train the prediction model on multiple nodes. The prediction model structure adopts a deep learning architecture combining LSTM and attention mechanism. Model compression deployment involves deploying models to edge computing devices through network pruning and parameter quantization. Online forecasting: A forecasting task is triggered every hour, taking in cleaned raw data from the past 168 hours to generate a forecast sequence for the next 24 hours. Interpretable outputs are analyzed using a shape-additive interpretation method to assess the factor contribution of each predicted value. Adaptive fine-tuning is triggered by the mean absolute percentage error (MAPE) of the model.

5. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 1, characterized in that, The carbon emission reduction benefit assessment module includes: Carbon emission reduction calculation, used to calculate the system's daily cumulative carbon reduction, is calculated using the following formula: ； in, This indicates the cumulative daily carbon reduction. This represents the optimized photovoltaic power at time t. The carbon emission reduction factor representing photovoltaic power generation; This represents the water production at time t. The carbon emission factor representing seawater desalination; Investment cost calculation: Calculate the total investment cost of the system, including equipment purchase cost, installation and commissioning cost, and operation and maintenance cost; The investment intensity index is generated by calculating the carbon reduction investment intensity index based on the daily cumulative carbon reduction and the total investment cost.

6. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 5, characterized in that, The formula for calculating carbon reduction investment intensity by the investment intensity index generation unit is as follows: ； in, This indicates the cumulative daily carbon reduction. Indicates the total investment cost. This indicates the intensity of carbon reduction investment, expressed in tons of carbon dioxide per 10,000 yuan per year.

7. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 1, characterized in that, The intelligent early warning and decision support module includes: Threshold monitoring involves regularly scanning system operation data to identify discrepancies in carbon reduction, abnormal investment intensity, and abnormal price fluctuations. Optimization suggestion: Regenerate the scheduling scheme by calling the dynamic power allocation module based on the anomaly type; Notification push notifications will send early warning information and optimization suggestions to relevant personnel via SMS, email, and mobile push channels.

8. The typical low-carbon engineering carbon reduction investment intensity calculation system according to claim 7, characterized in that, The carbon reduction difference calculated by the threshold monitoring unit is: ； in, This indicates differences in carbon reduction. This indicates the actual carbon reduction. This indicates the planned carbon reduction amount.

9. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, enable the execution of a typical low-carbon engineering carbon reduction investment intensity calculation system as described in any one of claims 1-8.

Images