A graph database consists of a collection of data based on graph structure and a computation engine based on graph traversal. The graph database uses vertices and edges to represent the data model. Data attributes are stored in vertices and edges, and relationships between data are described by edges between vertices (Yuning and Bingbing, 2022). Querying a collection of data based on graph structure is also known as graph traversal or graph computation. In the data model of a graph database, each vertex or edge can be equipped with computational functions, which are both data storage units and dynamic computational units. Parallelized graph computation can be easily achieved by performing local computation based on each vertex in the graph independently and carrying out information exchange through edges.

Building a grid model based on a graph database is the basis for applying graph computation technology to power system analysis and computation. The CIM/E model is a specification for describing power system models proposed by the State Grid Corporation of China (Yaozhong et al., 2006; Weimin et al., 2013). In this paper, we follow the CIM/E specification to build a grid model based on a graph database, and at the same time, we adapt the characteristics of the graph database to maximize the access efficiency of the model.

The graph model (GM) in a graph database can be represented as a ternary: G M = N , R , L , (1) where N = { n ₁, n ₂, n ₃, … , n _t } is a non-empty finite set of vertices and t elements in N denote t vertices; R = { r ₁, r ₂, r _3, … , r _s } is the edge set consisting of edges connecting ordered vertex even pairs (n _i , n _j ) in the vertex set N; s elements in R denote s edges and satisfy R ⊆ N × N; and L is the set of vertex attributes and edge attributes. Vertices and edges can have multiple attributes. Different types of vertices or edges are distinguished by their IDs. There can be several different edges between vertices, and the number of their attributes can be expanded dynamically.

Power system devices, such as buses, breakers, disconnectors, generators, and loads, are defined as vertices, and connection lines between devices are defined as edges in the graph database. In addition, in order to establish topological associations between devices, the physical node attributes describing the topological relationships between devices in the CIM/E model are defined as vertices. The correspondence between the grid entities and graph database elements is shown in Table 1, and an example of the grid model based on the graph database is shown in Figure 1.

The new energy data acquisition process for a certain range of new energy based on the graph database (taking the information of all new energy units connected through topological relationships under one main transformer as an example) contains the following steps:

1) In the given new energy plant station, select a main transformer and then find transformer’s medium-voltage side winding (labeled TransformerWinding2) and low-voltage side winding (labeled TransformerWinding3). These two windings of the same transformer are associated with each other by the same value of the transformer identification attribute.

The CLIQUE algorithm is a combination of density-based and mesh-based clustering methods and provides a new way of thinking about cluster analysis of high-dimensional data by introducing the concept of subspace clustering. The CLIQUE algorithm can explore the global distribution pattern of the data space by recognizing the density of cells in the data space. The central idea of the CLIQUE algorithm is as follows:

1) Given a large collection of multidimensional data points, which are usually not evenly distributed in the data space, the CLIQUE distinguishes between sparse and ‘crowded’ regions (or cells) in the space in order to discover the global distribution pattern of the data collection.

CLIQUE performs multidimensional clustering in two steps:

In the first step, CLIQUE divides the n-dimensional data space into mutually disjoint rectangular cells (Gengsheng et al., 2008) and identifies the dense cells in it. This work is performed for each dimension. The intersecting subspaces representing dense cells form a candidate search space in which higher-dimensional dense cells may exist (Lu and Zhang, 2020).

In the second step, CLIQUE generates a minimized description for each cluster. For each cluster, it determines the largest region that covers connected dense cells and then determines the smallest coverage (Xiangxiang and Dengsheng, 2012).

Compared to other clustering algorithms, such as k-means clustering and DBSCAN clustering, the CLIQUE algorithm does not need to set the initial number of clusters and does not need to assume any standard data distribution. It scales linearly with the size of the input data and has good scalability when the dimensionality of the data increases. The CLIQUE algorithm is very effective for clustering high-dimensional data in large-scale databases. In addition, the time and space complexity of the algorithm are linear, and the calculation speed is faster. The CLIQUE algorithm is a grid-based clustering method. The classification boundary in the clustering result is a grid boundary, so the boundary is a flat straight line, which makes it easy to determine the coupling relationship between the new energy prediction power and the prediction error. Therefore, the CLIQUE algorithm is selected for clustering new energy historical data.

The process of analyzing new energy historical operation information based on the CLIQUE algorithm contains the following steps:

1) Take {P ₁ = P ^’ } and {P ₂ = P-P ^’ }, i.e., the historical prediction power and historical prediction error, as clustering data.

The CLIQUE algorithm is used for the new energy historical data to analyze the coupling relationship between the predicted power and prediction error. Based on a given new energy prediction power, the confidence interval of the new energy actual power at the same moment as the prediction power can be obtained according to the corresponding prediction error distribution interval, which can effectively quantify the uncertainty of the new energy output (Sicheng and Weiqing, 2023).

Based on the coupling relationship between the prediction power and prediction error obtained from the cluster analysis, the lower and upper limits of the corresponding new energy prediction error are obtained according to the range of prediction power intervals to which the given new energy prediction data belong. The lower limit of the new energy prediction error is denoted as ERR _gd , and the upper limit of the new energy prediction error is denoted as ERR _gu .

Based on the given new energy power prediction value, the corresponding lower limit of the new energy actual power confidence interval is denoted as P e d = P e + E R R g d , (2) where P _e is the given new energy power prediction value and ERR _gd is the lower limit of the new energy prediction error.

The upper limit of the new energy actual power confidence interval is denoted as P e u = P e + E R R g u , (3) where ERR _gu is the upper limit of the new energy prediction error.

Based on the lower and upper limits of the new energy actual power confidence interval, the new energy actual power confidence interval is expressed as [P _ed , P _eu ].

It should be noted that the above calculation step is based on the predicted power of new energy at a single moment to calculate the confidence interval of the actual power of new energy at the same moment. On this basis, for the new energy prediction data over a period of time, the predicted power at each time point can be used to calculate the new energy actual power confidence interval at the same time following the above calculation step, and the new energy actual power confidence interval in the target time period can be obtained by traversing all the time points based on the prediction data.

Combining the topology of a model system containing some new energy plant stations, the CIM/E data are converted into vertex and edge datasets and loaded into a graph database to build a corresponding physical node graph model of the power system. Based on this graph database test dataset, the above method is used to obtain information on all new energy units under a certain main transformer connected through topological relationships, as shown in Table 2.

In Table 2, the left side contains the main transformer winding ID as the starting point of the search path and the right side contains the ID of the new energy generator found. For this graph database test dataset, a part of the new energy generators are found in main transformer low-voltage side winding (ID: 117375065841997076), and the rest of the new energy units are found in main transformer medium-voltage side winding (ID: 117375065841997077).

In Table 3, based on the ID of the new energy generator found, the information of the new energy generator searched under the main transformer, such as generator type and generator output, is obtained by the SQL query statement.

Based on the new energy historical data in a region obtained from the graph database, the new energy historical prediction and measured data are analyzed using the above method, and the operational distribution characteristics of the new energy historical prediction power and prediction error are obtained through the CLIQUE clustering method, as shown in Figure 2.

Based on Figure 2, the distribution intervals of prediction errors for new energy sources at different prediction power levels can be summarized, as shown in Table 4.

According to Figure 2 and Table 4, the fluctuation range of the prediction error is the largest when the new energy prediction power is 500 kW–1,000 kW. The range of fluctuation of the prediction error is the smallest when the predicted power is > 2,500 kW.

The new energy historical data in two different regions are analyzed and compared using the CLIQUE algorithm to obtain the coupling relationship between the new energy prediction power and prediction error in different regions, as shown below.

As can be seen from Figure 3 and Figure 4, the maximum range of new energy prediction error fluctuation in region 1 is [-2,000, 1,250], and the maximum range of new energy prediction error fluctuation in region 2 is [-2,500, 1,750]. According to the clustering results, it can be found that the new energy characteristics of different regions are more different.

For the region to which the new energy data analyzed in Figure 2 belongs, based on the new energy prediction data within a given day’s time, the new energy actual power confidence interval within this day is calculated according to the corresponding prediction error distribution interval, as shown in Figure 5.

In Figure 5, the red line represents the upper limit of the actual power confidence interval, and the blue line represents the lower limit of the actual power confidence interval. It can be seen that using the data-driven new energy uncertainty analysis method proposed in this paper, the new energy actual power confidence interval in the target time period is obtained, and the impact of the prediction error caused by new energy uncertainty is taken into account, realizing the quantitative description of new energy output uncertainty.