Abstract: With the continuous improvement of the digitalization and intelligence level of the energy supply and consumption system in agricultural parks, multi-type energy consumption data and multi-source heterogeneous data have shown an explosive growth trend. The complexity and correlation of data dimensions have increased significantly, which not only brings many complex challenges to dispatchers' reasoning, judgment, and decision-making but also exposes traditional modeling and optimization algorithms to shortcomings such as long strategy generation time and difficulty in quickly responding to dispatch needs when facing dynamically changing operating scenarios, seriously restricting the full play of the system's intelligent regulation efficiency. To address this, this paper proposes a knowledge graph-based optimization strategy retrieval method for the energy supply and consumption system in agricultural parks, aiming to solve the core problems of multi-source data processing and efficient strategy matching through the integrated application of knowledge engineering and deep learning technologies. Firstly, focusing on photovoltaic prediction data and load prediction data, a comprehensive structured data feature extraction system is constructed. This system includes both descriptive features such as meteorology and seasons, which can accurately characterize the differences in energy consumption scenarios under different environmental conditions, and digital features such as temporal change rate, mean value, peak-valley difference, and daily load rate, which comprehensively capture the dynamic change laws of data and overall operating characteristics. Through interval processing, digital features are converted into a standardized data feature label library, which effectively reduces redundant information in the knowledge graph and lays a solid data foundation for the subsequent construction of the knowledge graph. Secondly, knowledge extraction is carried out relying on deep learning technology. The Bi-LSTM-CRF model is used to accurately identify 21 types of named entities, including energy consumption scenarios, energy equipment, energy networks, and dispatch strategies, solving the problems of ambiguous boundaries and inaccurate classification of traditional methods in domain entity recognition. The improved Bi-GRU-Attention model introduces a dedicated terminology library and relationship priority coefficients for the agricultural energy supply and consumption field, assigning higher weights to core associative relationships, which further improves the accuracy and pertinence of mining associative relationships between entities and provides high-quality support for the association construction of the knowledge graph. On this basis, a knowledge graph of the energy supply and consumption system in agricultural parks, including a schema layer and a data layer, is constructed using the Neo4j graph database. The schema layer clarifies 6 types of core information such as energy consumption scenarios, energy equipment, operating data, and dispatch strategies, as well as corresponding associative relationships, forming a clear knowledge organization structure. The data layer effectively integrates the conversion results of structured data, semi-structured text, and unstructured images/videos through operations such as structured processing, feature extraction, and labeling, realizing the structured storage, visual display, and interactive query of historical strategy information and multi-dimensional data features. Furthermore, a hierarchical retrieval strategy is proposed. Through three core processes—information parsing, strategy feature matching, and dispatch strategy retrieval—feature extraction and label matching are first performed on the prediction data of the scenario to be optimized, then a knowledge search engine is used to find matching knowledge paths in the graph, and finally the optimal historical strategy is located. If no suitable strategy is matched, the optimization algorithm is activated to solve the new strategy, which is then updated to the graph and strategy library in a timely manner to realize the dynamic enrichment of the knowledge graph. Simulation example verification results show that in the experiment based on the typical winter day data of a northern agricultural park, compared with traditional modeling and optimization algorithms, this method shortens the strategy generation time from 22.5 s to 2.16 s, with an efficiency improvement of about 90%. Core optimization indicators such as the operating cost of 15,338.07 yuan and the useful energy conversion efficiency of 54.19% have a deviation of less than 1% from traditional methods. At the same time, it can intuitively display the multi-dimensional features of the operating scenario and the strategy matching path in the form of a subgraph, enhancing the readability and interpretability of dispatch results. This method effectively achieves efficient retrieval of optimization strategies and intelligent decision support for the energy supply and consumption system in agricultural parks. It not only ensures the reliability of optimization results but also significantly improves the dispatch response speed, providing a practical and feasible solution for the application of knowledge graphs in the field of integrated energy system dispatch and having important significance for promoting the intelligent upgrading of energy management in agricultural parks.