A Deep Learning-Based Method and System for Natural Resource Regulation

Abstract

This invention provides a deep learning-based method and system for natural resource regulation, relating to the field of natural resource regulation technology. The method includes: acquiring multi-source heterogeneous data; extracting dynamic features of resource supply and demand and spatial distribution features of the multi-source heterogeneous data; extracting deep temporal features of the multi-source heterogeneous data; fusing the deep temporal features, dynamic features of resource supply and demand, and spatial distribution features of resources to obtain fused features; predicting the future spatiotemporal state of natural resources based on the fused features using a multi-scale spatiotemporal prediction model; and regulating the natural resources based on the future spatiotemporal state using the AC-RL algorithm. This invention can significantly improve the predictability, accuracy, and robustness of natural resource regulation, effectively enhancing resource allocation efficiency and risk response capabilities.

Description

Technical Field

[0001] This invention relates to the field of natural resource regulation technology, and in particular to a natural resource regulation method and system based on deep learning. Background Technology

[0002] Deep learning-based natural resource regulation methods are a type of intelligent and refined management of natural resources (such as water, land, forests, and minerals) using deep learning technology. They typically utilize deep neural networks to automatically learn and represent the spatiotemporal evolution patterns, supply and demand trends, and external disturbance mechanisms of natural resources, thereby enabling the prediction, evaluation, and optimized regulation of resource status.

[0003] With the intensification of global climate change, the increasing frequency of human activities and the continuous development of the economy, the uneven spatial and temporal distribution of natural resources and the contradiction between supply and demand have become more and more prominent. The extensive management methods that rely on human experience and static models are no longer able to cope with the complexity and uncertainty of the system.

[0004] However, traditional natural resource regulation methods generally rely solely on single time series analysis or static spatial statistics, resulting in delayed identification of resource anomalies and insufficient prediction accuracy, failing to provide a reliable basis for regulation strategies. At the same time, most regulation strategies rely on human experience or static rules, lacking the ability to predict future spatiotemporal evolution and adaptive decision-making capabilities, which can easily lead to delayed resource regulation, inefficient allocation, or even risk accumulation. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a natural resource regulation method based on deep learning, which can solve the problems of existing natural resource regulation methods that generally rely only on single time series analysis or static spatial statistics, resulting in lagging identification of resource anomalies, insufficient prediction accuracy, inability to provide a reliable basis for regulation strategies, and regulation strategies that mostly rely on human experience or static rules, lacking the ability to predict future spatiotemporal evolution and adaptive decision-making ability, thus easily causing resource regulation lag, inefficient allocation, or even risk accumulation.

[0006] A first aspect of this invention proposes a natural resource regulation method based on deep learning, comprising:

[0007] S1: Acquire multi-source heterogeneous data;

[0008] S2: Extract the dynamic characteristics of resource supply and demand and the spatial distribution characteristics of resources from the multi-source heterogeneous data;

[0009] S3: Extract deep temporal features of the multi-source heterogeneous data using temporal convolutional networks and recurrent neural networks;

[0010] S4: Integrate the deep temporal features, resource supply and demand dynamic features, and resource spatial distribution features to obtain the fused features;

[0011] S5: Based on the fusion features, predict the future spatiotemporal state of natural resources using a multi-scale spatiotemporal prediction model;

[0012] S6: Based on the future spatiotemporal state, regulate the natural resources using the AC-RL algorithm.

[0013] A second aspect of the present invention provides a natural resource regulation system based on deep learning, comprising: a processor and a memory;

[0014] The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the deep learning-based natural resource regulation method as described in the first aspect.

[0015] A third aspect of the present invention provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the deep learning-based natural resource regulation method described in the first aspect.

[0016] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0017] In this embodiment of the invention, by integrating the dynamic characteristics of resource supply and demand, spatial distribution characteristics, and deep temporal characteristics, a comprehensive representation that can simultaneously characterize the temporal evolution law and spatial structural differences is constructed, realizing multi-dimensional modeling of the driving force of resource changes. Combined with a multi-scale spatiotemporal prediction model, it can make high-precision and fine-grained predictions of the future spatiotemporal state of natural resources, thereby identifying possible resource anomalies and risk trends in advance. By introducing an intelligent regulation mechanism based on the AC-RL algorithm, the regulation strategy can be adaptively optimized according to the prediction results, realizing dynamic closed-loop control of resource regulation. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart illustrating a deep learning-based natural resource regulation method provided in an embodiment of the present invention.

[0020] Figure 2This is a schematic diagram of a natural resource regulation system based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, in conjunction with the accompanying drawings, details the deep learning-based natural resource regulation method provided by the embodiments of the present invention through specific examples and application scenarios.

[0023] Reference manual attached Figure 1 The diagram illustrates a flowchart of a deep learning-based natural resource regulation method provided by an embodiment of the present invention.

[0024] This invention provides a deep learning-based method for regulating natural resources, which may include the following steps:

[0025] S1: Acquire multi-source heterogeneous data.

[0026] Specifically, multi-source heterogeneous data is acquired from data collection terminals of different sources, types, and time scales. These multi-source heterogeneous data include remote sensing images, meteorological observations, geographic information, resource supply records, resource demand records, economic and social statistical data, and policy regulation data.

[0027] S2: Extract the dynamic characteristics of resource supply and demand and the spatial distribution characteristics of resources from multi-source heterogeneous data.

[0028] Among them, the dynamic characteristics of resource supply and demand refer to the dynamic changes in natural resources over time, such as short-term change gradients, long-term trends, and periodic fluctuations, reflecting the evolutionary characteristics of resource supply and demand over time. The spatial distribution characteristics of resources refer to the distribution patterns and spatial structure of natural resources in different spatial locations, including spatial neighborhood statistics, autocorrelation coefficients, and spatial gradient strength, used to characterize the aggregation, diffusion, and abrupt changes of resources in space.

[0029] It should be noted that by simultaneously extracting the dynamic characteristics of resource supply and demand and the spatial distribution characteristics of resources, the laws governing changes in natural resources can be comprehensively depicted in both time and space dimensions, enabling a more accurate identification of the spatiotemporal evolution trend of resource status.

[0030] In one possible implementation, S2 specifically includes:

[0031] S201: Divide multi-source heterogeneous data into time dimensions and extract the resource supply, resource demand and resource change rate for each time period to form a resource supply and demand time series.

[0032] In this embodiment of the invention, the collected multi-source heterogeneous data is uniformly organized according to timestamps and divided into time dimensions according to a preset time granularity (such as hour, day, or ten-day period) to align data from different sources on the same time axis. Subsequently, key indicators such as resource supply, resource demand, and their rate of change are calculated within each time period, thereby forming a resource supply and demand time series that reflects the evolution of resource supply and demand over time. Resource supply and demand can be obtained by summing, averaging, or weighting the effective observations from corresponding data sources within each time period. For example, water withdrawal statistics can be used for water resources, and output and consumption records can be used for energy resources. After obtaining the supply and demand, the rate of change of resources can be obtained by calculating the difference or ratio between values ​​in adjacent time periods, reflecting the growth, decline, or fluctuation trend of resource supply and demand over continuous time.

[0033] S202: Based on the time series of resource supply and demand, calculate the short-term change gradient, long-term trend term, and periodic fluctuation term to obtain the dynamic characteristics of resource supply and demand.

[0034] The short-term gradient refers to the rapid magnitude and direction of changes in resource supply and demand within adjacent or recent time periods. It can be obtained through methods such as differencing and local slope estimation of time series data, and is used to reflect the short-term fluctuation characteristics of resource supply and demand. The long-term trend term refers to the overall direction and evolution of resource supply and demand over a longer time scale. It can be extracted using methods such as moving average regression, low-frequency filtering, or exponential smoothing, and is used to characterize the medium- to long-term growth, decline, or stabilization trends of resource supply and demand. The periodic fluctuation term refers to the repetitive fluctuation characteristics of resource supply and demand caused by external factors such as seasonal, rhythmic demand changes, or policy cycles. It is usually extracted through Fourier transform, wavelet decomposition, or periodic fitting models, and is used to describe the periodic structural patterns in resource changes.

[0035] Specifically, multi-scale differencing is performed on the resource supply and demand time series. Short-term gradients reflecting the rapid fluctuations in resource supply and demand are extracted by analyzing the slope changes of the most recent time steps. Subsequently, long-term trend analysis is performed on the time series using methods such as moving regression, exponential smoothing, or low-frequency filtering to obtain long-term trend terms describing the direction and persistence of resource evolution. By performing Fourier analysis, wavelet decomposition, or periodic template fitting on the time series, periodic fluctuation terms of resource supply and demand caused by seasonal and cyclical policy adjustments are extracted.

[0036] In this embodiment of the invention, by combining features from three scales—short-term gradient, long-term trend, and cyclical fluctuation—a dynamic resource supply and demand feature that comprehensively reflects the changing patterns of resource supply and demand over time can be constructed. , It can represent the short-term gradient. It can represent a long-term trend. It can represent the amplitude of the periodic term, and can be expanded into different resources, scales, and indicators. k dimension.

[0037] S203: Based on the distribution of natural resources in different spatial locations, multi-source heterogeneous data are mapped according to spatial grids to form a spatial resource distribution matrix.

[0038] S204: Calculate the spatial correlation of the spatial resource distribution matrix to obtain the spatial distribution characteristics of resources. Specifically, the spatial correlation includes: spatial neighborhood statistics, spatial autocorrelation coefficient, and spatial gradient strength.

[0039] Spatial neighborhood statistics refer to the characteristics obtained by statistical analysis of resource values ​​within each spatial grid location and its neighborhood, including indicators such as local mean, local variance, and neighborhood range. These are used to characterize the clustering, dispersion, and local stability of resources within a local spatial range. Spatial autocorrelation coefficient is an indicator that measures whether resources exhibit a correlated structure in space. By comparing the similarity of resource values ​​at adjacent or related locations, it reflects whether resources are clustered, dispersed, or randomly distributed on a larger spatial scale. Spatial gradient strength refers to the magnitude and direction of changes in resource values ​​between neighboring spatial locations. It is obtained by calculating pixel differences or spatial gradient operators (such as Sobel and Prewitt operators) and is used to describe abrupt boundaries, steep transition zones, and the degree of spatial gradient change in resources.

[0040] In this embodiment of the invention, multi-source heterogeneous information is spatially aligned and interpolated according to a unified grid coordinate system, so that each spatial location ( i , j A cell corresponds to a resource quantity or resource attribute, thus constructing a time frame. t The spatial resource distribution matrix is ​​then constructed. Subsequently, spatial neighborhood statistics such as local mean and local variance are calculated for each pixel and its neighborhood to characterize the clustering and dispersion characteristics of resources within a local area. Spatial autocorrelation coefficients such as Moran's I and Geary's C are used to measure whether there is a positive or negative spatial correlation structure of resources at a larger scale. The spatial gradient intensity obtained by differencing between pixels characterizes the significance of abrupt changes, boundary variations, and gradient changes in resources in space. By combining the above neighborhood statistics, autocorrelation, and gradient characteristics, a comprehensive spatial distribution feature of resources that can describe the spatial distribution pattern and spatial structure characteristics can be obtained.

[0041] In one possible implementation, the process after S2 and before S3 includes:

[0042] Based on the dynamic characteristics of resource supply and demand and the spatial distribution characteristics of resources, the natural resource anomaly degree is calculated, which includes the spatial distribution anomaly degree of resources and the anomaly degree of resource supply and demand.

[0043] In this embodiment of the invention, based on the dynamic characteristics of resource supply and demand, the standardized deviation of each supply and demand characteristic at the current moment relative to its historical mean and standard deviation, as well as the standardized trend deviation of each characteristic's rate of change relative to the historical rate of change distribution, are calculated to obtain the resource supply and demand anomaly degree characterizing the anomaly of supply and demand in the time dimension. Subsequently, based on the spatial resource distribution characteristics, by comparing the standardized differences in amplitude and fluctuation range between the resource values ​​at each spatial location and their historical spatial baseline, and combining this with the degree of local structural abrupt change reflected by the spatial gradient intensity, the resource spatial distribution anomaly degree characterizing the anomaly of the spatial pattern is obtained.

[0044] The specific formula for calculating the anomaly degree of natural resources is as follows:

[0045]

[0046] in, Z t express t The degree of natural resource anomaly at any given time. S t express t The anomaly of resource spatial distribution at any given time. This indicates the weight corresponding to the degree of anomaly in resource supply and demand. The weights represent the degree of spatial anomalies in resource distribution. D t express t The degree of anomaly in resource supply and demand at any given moment. K This represents the total number of dynamic characteristics of supply and demand. Indicates the first k Individual supply and demand dynamic characteristics t The value at time, Indicates the first k Historical averages of supply and demand dynamic characteristics Indicates the first k Historical standard deviation of supply and demand dynamic characteristics Indicates the first k The weighting coefficient of the magnitude deviation term of each supply and demand dynamic characteristic. Indicates the first k Weighting coefficients for abnormal trends in supply and demand dynamics. Indicates the first k Individual supply and demand dynamic characteristics trate of change at time, Indicates the first k The historical average of the rate of change of dynamic characteristics of supply and demand. Indicates the first k Historical standard deviation of the rate of change of dynamic characteristics of supply and demand. Indicates in t Time and spatial location ( i , j Resource characteristic observations, Indicates spatial location ( i , j The historical benchmark mean, Indicates spatial location ( i , j The historical standard deviation of the fluctuation. Indicates spatial location ( i , j Importance weighting coefficients M This indicates the number of rows in the spatial grid. N Indicates the number of columns in the spatial grid. The amplification factor that is sensitive to spatial gradients is represented by... express t Time and space location ( i , j Spatial gradient anomaly components, spatial gradient anomaly components It can be obtained by standardizing the deviation between the spatial gradient intensity and its historical gradient baseline to highlight local spatial abrupt change regions.

[0047] in, , Δ Δ All of these are calculated based on statistical measures of the dynamic characteristics of resource supply and demand over historical time series. , Based on the spatial baseline statistics of resource spatial distribution characteristics at historical moments, specifically, the spatial characteristic value at the current moment is compared with its historical spatial baseline (mean and standard deviation) by standardization, and combined with the abnormal components of spatial gradient intensity, the degree of abnormality of resources in local areas, overall spatial structure and boundary abrupt changes can be quantified, thus forming the resource spatial distribution anomaly degree.

[0048] When the anomaly degree of natural resources is greater than the threshold, proceed to step S3. When the anomaly factor is less than or equal to the threshold, return to step S1.

[0049] It should be noted that those skilled in the art can set the threshold according to actual needs, and this invention does not limit it.

[0050] Specifically, by comparing abnormal factors with preset thresholds, an adaptive anomaly screening and triggering mechanism can be implemented in the early stages of data processing. Only when the resource status deviates significantly will the process enter the deep time series analysis process, thereby effectively reducing unnecessary computational overhead, lowering the model false trigger rate, and improving the efficiency and stability of the overall control process.

[0051] S3: Extract deep temporal features from multi-source heterogeneous data using temporal convolutional networks and recurrent neural networks.

[0052] Temporal convolutional networks (TCNNs) are deep temporal modeling structures based on dilated and causal convolutions. They extract features by sliding fixed-length convolutional kernels along the time axis, capturing short-term local patterns and multi-scale temporal dependencies. They are characterized by high parallel computation efficiency and stable gradients. Recurrent neural networks (RNNs) are neural network structures that model sequence dependencies through recursive updates of hidden states.

[0053] In one possible implementation, S3 specifically includes:

[0054] S301: Align and resample multi-source heterogeneous data according to timestamps to obtain multi-dimensional time series sequences.

[0055] S302: Perform logarithmic transformation on a multidimensional time series.

[0056] Logarithmic transformation refers to a type of preprocessing operation that uses a logarithmic function (such as log(x+1) or log(x)) to perform nonlinear transformations on the values ​​in a multidimensional time series to compress the data amplitude, reduce the influence of extreme values, and improve the stationarity of the series. Logarithmic transformation can reduce the scale difference of the original data and make the rate of change more reflective of relative change characteristics.

[0057] S303: Slice the transformed multidimensional time series to obtain a time series sample set consisting of multiple time segments (where each sample in the time series sample set corresponds to a multi-source feature sequence within a continuous time period).

[0058] S304: Input the time series sample set into the time series convolutional network to extract local time series feature representations associated with short-term changes.

[0059] S305: By using a recurrent neural network, the temporal dependency model of the local temporal feature representation is performed to obtain the deep temporal hidden state sequence.

[0060] S306: Aggregate the deep temporal hidden state sequence along the time dimension to obtain deep temporal features.

[0061] In this embodiment of the invention, the constructed temporal sample set is input into a temporal convolutional network. Local temporal features reflecting short-term fluctuations and local change patterns are extracted by sliding multi-scale convolutional kernels along the time axis. These local features are then input into a recurrent neural network, where long-term dependencies across time periods are captured through recursive hidden state updates, forming a temporal hidden state sequence capable of representing deep dynamic structures. Finally, the hidden state sequence is weighted and aggregated or pooled along the time dimension (e.g., using global average pooling or attention-weighted summation) to obtain deep temporal features that combine local change characteristics with long-term dependency information, providing high-quality temporal representation input for subsequent spatiotemporal prediction models.

[0062] S4: Integrate deep temporal features, dynamic features of resource supply and demand, and spatial distribution features of resources to obtain integrated features.

[0063] In this embodiment of the invention, the deep temporal features extracted by the time series model are aligned, normalized and multidimensionally fused with the dynamic features of resource supply and demand and the spatial distribution features of resources in the feature space. Through feature splicing, weighted superposition or nonlinear mapping, the temporal evolution law, supply and demand change pattern and spatial structure information are comprehensively encoded in a unified high-dimensional expression, so that the obtained fused features can simultaneously reflect the temporal trend, supply and demand status and spatial distribution characteristics of resources.

[0064] S5: Based on the fusion characteristics, predict the future spatiotemporal state of natural resources through a multi-scale spatiotemporal prediction model.

[0065] It should be noted that by processing the fused features through a multi-scale spatiotemporal prediction model, accurate modeling of the future evolution trend of natural resources can be achieved by simultaneously considering local detail changes and global spatial structure. The multi-scale spatiotemporal prediction model can capture spatiotemporal dependencies across scales, making the prediction results more continuous in the time dimension and more consistent in the spatial dimension, thereby significantly improving the accuracy and stability of inferring future spatiotemporal states.

[0066] In one possible implementation, the multi-scale spatiotemporal prediction model specifically includes: an input module, a fusion feature construction module, a multi-scale spatial feature extraction module, a spatiotemporal encoding module, a spatiotemporal decoding module, a state prediction module, a spatiotemporal mapping module, and an output module.

[0067] S5 specifically includes:

[0068] S501: In the input module, input fusion features.

[0069] S502: In the fusion feature construction module, the fusion features are organized into a multi-channel raster map according to the spatial grid to obtain a spatiotemporal feature map.

[0070] S503: In the multi-scale spatial feature extraction module, multi-scale feature extraction is performed on the spatiotemporal feature map to obtain a multi-scale spatial feature sequence.

[0071] In this embodiment of the invention, the constructed spatiotemporal feature map is input into the multi-scale spatial feature extraction module. By setting convolution kernels with different receptive fields, pyramid-shaped feature extraction structures, or multi-level spatial downsampling and upsampling paths, the feature map is analyzed at multiple scales. This can simultaneously capture local fine-grained information and large-scale spatial structure patterns, forming a multi-scale spatial feature sequence composed of spatial features at different scales.

[0072] S504: In the spatiotemporal coding module, the multi-scale spatial feature sequence is spatiotemporally encoded using a ConvLSTM encoder to obtain the encoder state. Specifically, the encoder state includes the decoder hidden state sequence and the decoder memory unit sequence. :

[0073]

[0074] in, express t Time input gate, This represents the Sigmoid activation function. This represents the input convolution kernel of the input gate. express t Multi-scale spatial feature sequence at time step The hidden state of the input gate is represented by the convolution kernel. express t- The hidden state at time 1 This represents the memory coupling weights of the input gate. express t- The memory state at moment 1 This indicates the input gate bias. express t The Gate of Forgetting Time This represents the input convolution kernel of the forget gate. The hidden state convolution kernel representing the forget gate, The memory coupling weights of the forget gate are represented. Indicates the forget gate bias. express t The state of memory at any given moment express Activation function The input convolution kernel represents the candidate memory. Hidden convolutional kernels representing candidate memories, This indicates candidate memory bias. express t The output gate of the moment, This represents the input convolution kernel of the output gate. This represents the hidden convolutional kernel of the output gate. This represents the hidden convolutional kernel of the output gate. Indicates the output gate bias. This indicates element-wise multiplication. express t The hidden state at all times This represents a two-dimensional convolution.

[0075] Among them, the ConvLSTM (Convolutional Long Short-Term Memory) encoder is a spatiotemporal network model that embeds a convolutional structure inside a Long Short-Term Memory (LSTM) unit. By using convolutional operations to replace the fully connected operations of traditional LSTM, it can capture dynamic trends in the time dimension while preserving local structural features in the spatial dimension, thus achieving joint modeling of spatiotemporal sequence data.

[0076] In this embodiment of the invention, a multi-scale spatial feature sequence is input into a spatiotemporal coding module. A ConvLSTM encoder utilizes a convolutional structure and a memory gating mechanism to jointly model the sequence spatiotemporally, enabling the network to capture temporal dynamics while preserving local spatial structural information. During the encoding process, ConvLSTM generates hidden state sequences step-by-step. H t Used to characterize the spatiotemporal features of the current moment, while updating the memory unit sequence. C t It stores long-term dependency information across time periods, and together the two constitute the complete encoder state.

[0077] S505: In the spatiotemporal decoding module, the encoder state is spatiotemporally decoded to obtain the future fused feature sequence.

[0078] It should be noted that by decoding the encoder state step by step through the spatiotemporal decoding module, the feature representation of future moments can be gradually reconstructed while maintaining the continuity of long-term memory information and spatial structure, thereby generating a future fused feature sequence. This can effectively transmit the spatiotemporal dependencies captured in the encoding stage, making the prediction results smoother in the time dimension and more consistent in the spatial dimension, and significantly improving the accuracy and stability of the overall future state inference.

[0079] In one possible implementation, S505 specifically includes:

[0080] S5051: Weight the decoder hidden state sequence and the decoder memory unit sequence to obtain the seed hidden state sequence and the seed memory unit sequence.

[0081] S5052: Construct a spacetime drift field.

[0082] In this embodiment of the invention, the hidden state of the encoder at the final moment of the input sequence is used as the reference. As the basic input for drift information, a drift vector field for each spatial location is generated through convolution mapping. ,in and They represent spatial locations respectively Future characteristics along x direction and y directional migration components, Indicates the drift field convolution kernel. This represents the drift field bias. A spatiotemporal drift field is constructed to characterize the main spatial migration dynamics of natural resource features in the future time dimension. Specifically, the encoder's hidden state at the current moment is used as the basic input for drift information. A two-dimensional drift vector is generated for each spatial location through convolutional mapping, thus obtaining a spatiotemporal drift field reflecting local change trends. The drift field contains drift components in the horizontal and vertical directions, which can be used to describe the intensity and direction of future feature movement in space, providing dynamic constraints for subsequent semi-Lagrange pre-updates and achieving physical consistency propagation of future features along the dominant motion direction.

[0083] S5053: Based on the spatiotemporal drift field, perform a semi-Lagrange pre-update on the seed hidden state sequence to obtain the drift hidden state sequence. .

[0084] The formula for the semi-Lagrange pre-update is as follows:

[0085]

[0086] in, express Time seed hidden state sequence in x Advection displacement in the direction of flow Indicates the time index of the prediction phase. Represents the drift field x Quantity, express Time seed hidden state sequence in y Advection displacement in the direction of flow Represents the drift field y Quantity, express The drifting hidden state sequence at time points, L ( ) represents a spatial interpolation function. express The hidden state of the seed at any given moment. T out Indicates the length of the future output time series. This indicates the length of the historical input time, that is, the last moment of the historical time-series data received by the encoder. This indicates the duration of the future prediction sequence, i.e., the number of future time steps that need to be predicted.

[0087] In this embodiment of the invention, for the seed hidden state sequence generated at the end of the encoder, the horizontal and vertical components of the spatiotemporal drift field are used to calculate the advection displacement at each future time step τ, and the seed hidden state is projected to the future spatial position in the manner of "current position minus translation amount". Then, the translated position is valued by a spatial interpolation function to generate the drift hidden state sequence. The semi-Lagrange pre-update process can simulate the continuous evolution of features along the dominant spatial migration direction over time, making the prediction of future spatiotemporal features more consistent with the translation, diffusion and dynamic flow of natural resources in real space.

[0088] S5054: Perform structural enhancement on the drift hidden state sequence.

[0089] For example, structural enhancement can be achieved by introducing residual convolution modules, attention enhancement modules, or multi-scale feature resampling modules to strengthen the boundary details and structural textures in the drifting hidden state sequence.

[0090] S5055: Based on the natural resource anomaly degree and combined with the enhanced drift hidden state sequence, gating correction is applied to the drift hidden state sequence and seed memory unit sequence to obtain the decoded hidden state sequence and decoded memory unit sequence:

[0091]

[0092] in, Indicates spatial location ( x , y The abnormal gating coefficient at ) Indicates the gating weight coefficient. b Indicates the gating bias term. Indicates in The spatial location at time is ( x , y Natural resource anomaly () The anomaly rate can be extrapolated based on the predicted features of future times and the historical baseline using the aforementioned anomaly rate calculation formula, thereby enabling gating correction of future anomaly regions during the decoding stage. Indicates in The spatial location at time is ( x , y Decoding the hidden state sequence of ) Indicates in The spatial location at time is ( x , yThe sequence of decoded memory units, Indicates in The spatial location at time is ( x , y The decoder hidden state sequence, Table The spatial location at time is ( x , y The seed memory unit sequence, Indicates in The spatial location at time is ( x , y The decoder memory unit sequence.

[0093] In this embodiment of the invention, a gating coefficient is constructed by the anomaly degree of natural resources, and the gating coefficient is combined with the enhanced drift hidden state sequence to perform weighted correction on the hidden state sequence obtained by future drift and the corresponding seed memory unit sequence. Specifically, for spatial locations with high anomaly degree, the gating coefficient will increase the weight of the historical baseline state in the correction process, thereby suppressing the disordered diffusion of future features in the anomaly region and generating a more robust decoded hidden state sequence and decoded memory unit sequence.

[0094] S5056: Perform a reversible channel transformation mapping on the decoded hidden state sequence and the decoded memory unit sequence to obtain the future fused feature sequence.

[0095] In embodiments of the present invention, future fusion features are generated by performing reversible channel transformation mapping on the decoded hidden state sequence and the decoded memory unit sequence. Specifically, the decoded hidden states are grouped according to the channel dimension. The first group of channels generates corresponding scale and translation parameters, and the second group of channels is reversibly mapped at the channel level based on an affine coupling layer. This allows the transformed features to achieve feature reconstruction and enhancement while maintaining information reversibility. Subsequently, the transformed groups of channels are reassembled, and a fusion feature map corresponding to the future time step is generated through convolutional mapping. By sequentially performing this reversible channel transformation and convolutional mapping operation on all future time steps, the future fusion feature sequence can be obtained, achieving stable generation and reversible reconstruction of future spatiotemporal features.

[0096] S506: Predict the state of natural resources based on future fusion feature sequences.

[0097] In this embodiment of the invention, by predicting the resource status based on future fusion feature sequences, it is possible to achieve high-precision judgment of future resource evolution trends by making full use of multi-scale spatial structure, deep temporal dependence and supply and demand dynamic information. It can not only capture complex change patterns across time, but also integrate multiple types of resource factors to improve the stability and reliability of prediction results.

[0098] In one possible implementation, S506 specifically includes:

[0099] S5061: Perform temporal context feature aggregation on the future fused feature sequence to obtain temporal context features.

[0100] S5062: Map the temporal context features to multiple resource factor subspaces to obtain multiple resource state representations.

[0101] In this embodiment of the invention, the decoded future fusion feature sequence is first aggregated with temporal context features. By constructing attention weights in the time dimension, dynamic correlation modeling is performed on the features at each future moment, thereby obtaining temporal context features that can represent cross-time dependencies. Subsequently, these temporal context features are mapped to multiple resource factor subspaces, and semantic components related to the state of natural resources are extracted through multi-factor weighted projection, including various types of feature expressions such as hydrological factors, ecological factors, land use factors, and climate factors, to construct a resource state representation that can comprehensively reflect the driving mechanism of resource changes, providing a high-dimensional structured input for subsequent state probability modeling and risk-sensitive prediction.

[0102] S5063: Based on the state representation of each resource, perform state probability modeling to obtain the state probability distribution:

[0103]

[0104] in, Indicates the first c Natural resources in the first At any given time, the spatial location is Energy value, c Indexes representing categories of natural resource status (such as land use type, water resource level, ecological risk level, etc.). w c Indicates the first c First-order linear weight vector of natural resources Indicates the first At any given time, the spatial location is Resource status representation, Indicates the first c The second-order interaction matrix (symmetric matrix) of natural resources is used to describe The quadratic relationship between the characteristic components b c Indicates the first c Energy bias term for natural resources, Indicates the first Time, spatial location This place belongs to the first cThe predicted probability of natural resources, where exp represents the exponential function. Indicates the temperature coefficient. C This represents the total number of natural resource status categories. T Indicates transpose. Indicates the first k Natural resources in the first At any given time, the spatial location is The energy value.

[0105] S5064: Combining risk-sensitive scoring functions and state probability distributions to predict the state of natural resources:

[0106]

[0107] in, Indicates the first c Natural resources in the first At any given time, the spatial location is Risk sensitivity score Represents the probability sensitivity coefficient. Indicates the first c Natural resources in the first At any given time, the spatial location is Risk indicators (such as resource depletion risk, ecological damage risk, land degradation risk, etc.). Indicates the risk penalty coefficient. Indicates the first At any given time, the spatial location is The state of natural resources is the state category, i.e., the state prediction result.

[0108] S507: In the output module, the state prediction results are mapped to obtain the future spatiotemporal state.

[0109] Specifically, the natural resource state prediction results obtained through risk-sensitive prediction are input into the output module. A spatial mapping function maps this state from the feature space or category space to the corresponding physical quantity or classification label in the actual geographic grid. More specifically, based on the temporal sequence structure and spatial location index of the state prediction results, the state values ​​at each future time point are restored pixel-by-pixel, interpolated, or recalibrated, transforming them into a two-dimensional or three-dimensional spatiotemporal distribution map with real geographic meaning. This generates a future spatiotemporal state representing the future spatiotemporal changes of natural resources.

[0110] In this embodiment of the invention, seed states are constructed using the hidden states output by the encoder and memory units, and a spatiotemporal drift field describing the future feature migration trend is generated based on convolutional mapping. This enables the network to simulate the dynamic migration of spatial information in a physically consistent manner. Subsequently, through a semi-Lagrange pre-update and structure enhancement mechanism, combined with anomaly factors for gating correction, a stable and structurally continuous future fusion feature sequence is generated step by step. Through temporal context aggregation, factor subspace mapping, energy model probability inference, and risk-sensitive scoring, a fine prediction of the future state of natural resources is achieved. Finally, the prediction results are mapped onto the actual spatial grid to form the future spatiotemporal state, thereby achieving a high-precision, interpretable, and physically consistent prediction output of future changes in natural resources.

[0111] S6: Based on future spatiotemporal conditions, natural resources are regulated using the AC-RL algorithm.

[0112] It should be noted that by performing resource regulation through the AC-RL algorithm, the regulation strategy can achieve adaptive optimization based on the prediction results, dynamically balance resource supply and demand, risk level and regulation cost, and continuously improve decision quality in complex environments, thereby significantly improving the intelligence, response efficiency and global optimization level of natural resource regulation.

[0113] In one possible implementation, S6 specifically includes:

[0114] S601: Set the state space according to the future spacetime state.

[0115] The state space describes the overall state of natural resources across multiple time scales, including their evolutionary trends, risk levels, and external environmental factors. It is denoted as... , This represents the vector of external influencing factors at the current moment, such as climate pressure, policy control restrictions, and environmental disturbance information. This represents the boundary information of current controllable factors of natural resources, such as resource supply capacity, spatial controllable range, and maximum adjustment range. The state space can continuously reflect future spatiotemporal prediction results and risk assessment information, enabling reinforcement learning agents to perceive the current state of natural resources and future spatiotemporal evolution trends when making control decisions.

[0116] S602: Define the action space and reward function based on the state space.

[0117] Specifically, the scope of action includes: adjusting the supply of resources (such as increasing / decreasing the allocation of water resources), adjusting the land use structure (increasing / decreasing the area of ​​a certain type of region), adjusting the intensity of ecological protection (increasing / decreasing the intensity of a certain type of protection measures), and implementing resource restoration measures (activating / stopping a certain type of restoration operation).

[0118] The reward function is as follows:

[0119]

[0120] in, express t The reward function at time step, express t The cost of resource deviation under the constant natural resource state. express t The risks and costs of natural resource status at all times. Indicates the state of natural resources Next, predict the first Time and space The resource state representation vector, Indicates spatial location The corresponding target resource status representation, Let |||2 denote the set of spatial grids, and |||2 denote the L2 norm. Indicates the first Weighting coefficients for risks similar to natural resources Indicates the state of natural resources Below, the predicted first Time and space Belongs to the The probability of a state similar to that of natural resources.

[0121] S603: Construct a natural resource reinforcement learning environment model based on the state space, action space, and reward function.

[0122] S604: Based on a natural resource reinforcement learning environment model, natural resources are regulated through the AC-RL algorithm.

[0123] In one possible implementation, the AC-RL algorithm specifically includes a critic policy network and an actor policy network.

[0124] S604 specifically includes:

[0125] S604A: Initialize the parameters of the critic policy network, the actor policy network, and the target policy network.

[0126] S604B: Based on the state of the actor's policy network, the action is sampled to obtain the sampled action.

[0127] S604C: A reinforcement learning environment model based on natural resources, which determines the next state and reward based on the sampled action.

[0128] S604D: Stores the state, sampled action, next state, and reward of the actor's strategy network into the experience pool.

[0129] S604E: Randomly extract historical data from the experience pool.

[0130] S604F: Update the critic strategy network based on historical data.

[0131] S604F specifically includes:

[0132] Calculate the Bellman error for historical data.

[0133] Update the commentator's policy network based on the Bellman error:

[0134]

[0135] in, Representing historical data e Bellman error, Indicates the first e Historical data, Indicates the state The corresponding resource deviation cost, Representing state The corresponding risk and cost, Indicates that the actor network is in a state The output action, Represents the action weight matrix. Represents the gradient. This represents the value function as characterized by the critic policy network. This represents the network parameters of the critic strategy network. Indicates the state Below, the predicted values ​​of state changes or the estimated results of state transition functions given by environmental dynamics approximation models (such as multilayer perceptrons) are... The parameters represent the approximate model of environmental dynamics. Represents the control gain matrix. This indicates the updated critic strategy network. Indicates the online learning rate of critics. This represents the least-squares adaptive gain matrix of the critic network. Indicates the current time t The weighted vector of the sampled data, Indicates the current time t The weighting coefficients of the sampled data, Representing historical data e The weight, N This represents the total number of historical data. Representing historical data e The weighted vector, Indicates the current time tThe normalization factor of the sampled data, Representing historical data e normalization factor ( t This represents the real-time sampled data corresponding to the current moment, while e (This represents the historical data stored in the experience pool).

[0136] S604G: Based on feedback from the updated critic strategy network, the parameters of the actor strategy network are updated:

[0137]

[0138] in, This represents the updated actor strategy network parameters. This represents the network parameters of the actor strategy. Indicates the actor's online learning rate. This represents the first sample obtained from the experience pool. Historical data, Indicates the state Take action At that time, the critic strategy network uses parameters The calculated action value function (Q value). This represents the network parameters for the actor strategy. gradient operator, This indicates that the actor policy network is in a given state. Below, based on its current parameters The output action, This indicates that the online commentators' opinions on the action a The partial derivative gradient, N This represents the total number of historical data.

[0139] S604H: Based on the updated actor policy network, update the parameters of the target policy network to obtain the control policy.

[0140] S604J: Determine if the maximum number of iterations has been reached. If yes, output the control strategy as the optimal control strategy. Otherwise, return to step S604B.

[0141] In this embodiment of the invention, the updated actor policy network is used as a reference model to perform soft updates on the parameters of the target policy network, enabling the target policy network to gradually approach the current optimal policy while maintaining parameter stability. Specifically, in each iteration, the target policy network smoothly adjusts its parameters according to a preset update coefficient, bringing it closer to the parameter direction of the actor policy network, thereby generating a stable control policy.

[0142] Reference manual attached Figure 2The diagram shows a schematic representation of a deep learning-based natural resource regulation system provided in an embodiment of the present invention.

[0143] This invention provides a deep learning-based natural resource regulation system 20, comprising: a processor 201 and a memory 202;

[0144] The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the deep learning-based natural resource regulation method described above and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0145] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0146] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0147] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0148] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0149] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0150] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0151] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0154] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] This invention provides a readable storage medium comprising: storing a program or instructions on the readable storage medium, wherein when the program or instructions are executed by a processor, the program or instructions implement the steps of the above-described deep learning-based natural resource regulation method and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A natural resource regulation method based on deep learning, characterized in that, The method comprises the following steps: S1: acquiring multi-source heterogeneous data; S2: extracting resource supply and demand dynamic characteristics and resource spatial distribution characteristics of the multi-source heterogeneous data; wherein, after the S2 and before the S3, further comprising: calculating a natural resource anomaly degree according to the resource supply and demand dynamic characteristics and the resource spatial distribution characteristics, wherein the natural resource anomaly degree specifically comprises a resource spatial distribution anomaly degree and a resource supply and demand anomaly degree; when the natural resource anomaly degree is greater than a threshold, entering step S3; when the natural resource anomaly degree is less than or equal to the threshold, returning to step S1; S3: extracting deep time sequence characteristics of the multi-source heterogeneous data through a time sequence convolution network and a recurrent neural network; S4: fusing the deep time sequence characteristics, the resource supply and demand dynamic characteristics, and the resource spatial distribution characteristics to obtain fused characteristics; S5: predicting a future spatio-temporal state of a natural resource through a multi-scale spatio-temporal prediction model according to the fused characteristics; wherein, the multi-scale spatio-temporal prediction model specifically comprises an input module, a fused characteristic construction module, a multi-scale spatial feature extraction module, a spatio-temporal encoding module, a spatio-temporal decoding module, a state prediction module, a spatio-temporal mapping module, and an output module; the S5 specifically comprises: S501: inputting the fused characteristics in the input module; S502: organizing the fused characteristics into a multi-channel raster map according to a spatial grid in the fused characteristic construction module to obtain a spatio-temporal feature map; S503: performing multi-scale feature extraction on the spatio-temporal feature map in the multi-scale spatial feature extraction module to obtain a multi-scale spatial feature sequence; S504: performing spatio-temporal encoding on the multi-scale spatial feature sequence through a ConvLSTM encoder in the spatio-temporal encoding module to obtain an encoder state, wherein the encoder state specifically comprises a decoder hidden state sequence and a decoder memory cell sequence: S505: performing spatio-temporal decoding on the encoder state in the spatio-temporal decoding module to obtain a future fused characteristic sequence; the S505 specifically comprises: S5051: weighting the decoder hidden state sequence and the decoder memory cell sequence to obtain a seed hidden state sequence and a seed memory cell sequence; S5052: constructing a spatio-temporal drift field; S5053: performing semi-Lagrangian pre-update on the seed hidden state sequence according to the spatio-temporal drift field to obtain a drift hidden state sequence; S5054: performing structure enhancement on the drift hidden state sequence; S5055: based on the natural resource anomaly degree, combining the enhanced drift hidden state sequence to perform gated correction on the drift hidden state sequence and the seed memory cell sequence to obtain a decoding hidden state sequence and a decoding memory cell sequence; S5056: performing reversible channel transformation mapping on the decoding hidden state sequence and the decoding memory cell sequence to obtain the future fused characteristic sequence; S506: predicting a state of the natural resource based on the future fused characteristic sequence; wherein, the S506 specifically comprises: S5061: performing temporal context feature aggregation on the future fusion feature sequence to obtain a temporal context feature; S5062: mapping the temporal context feature to a plurality of resource factor subspaces to obtain a plurality of resource state representations; S5063: based on each of the resource state representations, performing state probability modeling to obtain a state probability distribution; S5064: combining a risk sensitivity scoring function and the state probability distribution to perform state prediction on the natural resource; S507: mapping the state prediction result in the output module to obtain the future spatio-temporal state; S6: according to the future spatio-temporal state, regulating and controlling the natural resource through the AC-RL algorithm. 2.The deep learning-based natural resource regulation method of claim 1, wherein, The S2 specifically comprises: S201: dividing the multi-source heterogeneous data in the time dimension, and extracting the resource supply amount, resource demand amount and resource change rate of each time period to form a resource supply and demand time sequence; S202: based on the resource supply and demand time sequence, calculating a short-term change gradient, a long-term trend item and a periodic fluctuation item to obtain the resource supply and demand dynamic feature; S203: according to the distribution of the natural resource at different spatial positions, mapping the multi-source heterogeneous data according to the spatial grid to form a spatial resource distribution matrix; S204: calculating the spatial correlation of the spatial resource distribution matrix to obtain the resource spatial distribution feature, wherein the spatial correlation specifically comprises: spatial neighborhood statistics, spatial autocorrelation coefficient and spatial gradient intensity. 3.The deep learning-based natural resource regulation method of claim 1, wherein, The S3 specifically comprises: S301: aligning and resampling the multi-source heterogeneous data according to the timestamp to obtain a multi-dimensional time sequence; S302: performing logarithmic transformation processing on the multi-dimensional time sequence; S303: slicing the multi-dimensional time sequence after transformation processing to obtain a time sequence sample set composed of a plurality of time segments; S304: inputting the time sequence sample set into the time sequence convolution network to extract a local time sequence feature representation associated with short-term change; S305: modeling the time sequence dependence of the local time sequence feature representation through the recurrent neural network to obtain a deep layer time sequence hidden state sequence; S306: performing time dimension aggregation on the deep layer time sequence hidden state sequence to obtain the deep layer time sequence feature. 4.The deep learning-based natural resource regulation method of claim 1, wherein, The S6 specifically comprises: S601: according to the future spatio-temporal state, setting a state space; S602: according to the state space, defining an action space and a reward function; S603: according to the state space, the action space and the reward function, constructing a natural resource reinforcement learning environment model; S604: based on the natural resource reinforcement learning environment model, regulating and controlling the natural resource through the AC-RL algorithm.

5. The natural resource regulation method based on deep learning according to claim 4, characterized in that, The AC-RL algorithm specifically comprises: a critic policy network and an actor policy network; The S604 specifically comprises: S604A: initializing the parameters of the critic policy network, the parameters of the actor policy network and the parameters of the target policy network; S604B: based on the state of the actor policy network, sampling an action to obtain a sampled action; S604C: determining a next state and a reward according to the sampled action based on the natural resource reinforcement learning environment model; S604D: storing a state of the actor policy network, the sampled action, the next state and the reward into an experience pool; S604E: randomly extracting historical data from the experience pool; S604F: updating the critic policy network according to the historical data; S604G: updating parameters of the actor policy network according to feedback of the updated critic policy network; S604H: updating parameters of the target policy network according to the updated actor policy network to obtain a regulation policy; S604J: determining whether a maximum iteration number is reached; if yes, outputting the regulation policy as an optimal regulation policy; otherwise, returning to step S604B until the maximum iteration number is reached.

6. A readable storage medium characterized by, The program or instruction stored on the readable storage medium is executed by the processor to implement the steps of the deep learning-based natural resource regulation method according to any one of claims 1 to 5. The program or instruction stored on the readable storage medium is executed by the processor to implement the steps of the deep learning-based natural resource regulation method according to any one of claims 1 to 5.

Images