BiGRU-CRF system reaches 96.52% accuracy in power-grid text

A domain-specific artificial-intelligence system reached 96.52% entity-level exact-match accuracy when identifying information in Chinese power-grid text, according to a 2026 experimental study. The system combined a bidirectional gated recurrent unit, conditional random fields and natural-language-processing tools designed to extract entities from inspection records, dispatch logs, equipment-maintenance reports and public grid-company documents. Across 6,000 text samples, the model identified equipment, people and organizations, operational data, geographical locations and time expressions.

In an optimized test configuration, its reported average processing time was 28.35 milliseconds for 10,000 entities, a laboratory result that the authors said could support near-real-time analysis but would vary with hardware and data characteristics. The computational study addressed two practical challenges in power-system text: specialized terminology that can be segmented incorrectly and unequal frequencies among entity labels, which can make minority categories difficult to learn. It also sought to identify multiword entities with more precise boundaries.

The resulting pipeline had three stages: construction of a power-domain dictionary, probability-based word segmentation and neural sequence labeling with a BiGRU-CRF architecture. The dataset was divided by stratified sampling into 4,200 training items, 900 validation items and 900 test items, and the recognition model was trained on the 4,200-item training split rather than on the larger corpus used for language-model estimation. The dictionary was assembled from technical terms and abbreviations collected under national standards and from internal grid-enterprise data.

Each word was converted to a fixed-length MD5 hash and assigned to a bucket through a modulo operation. A chi-square test was used to examine the uniformity of the hash distribution, and the reported collision rate was below 0.5%. The researchers selected an improved cuckoo-hashing structure to balance collision handling, query speed and dynamic updating.

In a separate dictionary experiment on an Intel i7-12700H system with 16 GB of memory, the proposed structure took 98.7 milliseconds to construct, achieved an average query time of 0.9 microseconds and occupied 5.6 MB. The optimized Trie tree recorded 259.4 milliseconds, 2.8 microseconds and 8.9 MB, while the minimal perfect hash required 352.1 milliseconds, returned queries in 2.4 microseconds and occupied 12.3 MB. Forward maximum matching was then used to generate possible word divisions from the power dictionary.

A bigram probabilistic language model selected among those candidates by calculating the joint probability of each sequence. This language model, not the neural entity recognizer, was trained on an approximately 500,000-word power-domain corpus assembled from professional literature, equipment manuals and operation and maintenance reports. Kneser-Ney smoothing supplied probability estimates for unknown words and rare professional n-grams, while log probabilities were used to avoid numerical underflow during sequence scoring.

The resulting ordered word sequence, with position information and BIO labels, became the input to the recognition network. The core sequence model was a bidirectional GRU with an input dimension of 200 and a hidden dimension of 512. Separate forward and backward recurrent units allowed each word representation to incorporate preceding and following context.

The GRU was chosen as a lighter alternative to a bidirectional long short-term memory network, combining three gates into two and reducing the number of matrix parameters while preserving bidirectional sequence modeling. Its gates were described as adapting to state changes in power-equipment descriptions and filtering redundant historical information. Two further modifications addressed the structure of the annotated data.

The CRF training objective used inverse-frequency weights, giving less frequent tags greater influence than common tags and thereby addressing the observed label imbalance. A semi-Markov CRF was also introduced to model complete entity fragments, including their start, end and length, instead of decoding every label independently. The combined objective balanced standard and weighted loss terms with a regularization term, and grid search set the weighting coefficient to 10^-5 after selecting the best validation F1 score and loss balance.

In the ablation experiments, removing the weighted loss reduced macro-average F1 by 3.28 percentage points. Replacing the semi-Markov CRF with a traditional CRF produced the largest decrease among the ablations, which the authors interpreted as evidence that explicit fragment modeling contributed to boundary recognition for multiword entities. The labeled dataset contained 98,100 entities: 37,335 time entities, 32,409 equipment entities, 10,262 geographical entities, 9,238 person and organization entities and 8,856 operational-data entities.

A unified fine-grained scheme was applied across the text, and preprocessing included coding standardization, noise filtering, terminology unification, format standardization and sentence segmentation. Labels passed through initial annotation, cross-checking and expert audit. Reported inter-annotator consistency exceeded 0.85 for the entity groups and reached 0.92 for equipment entities.

The researchers distinguished entity-level exact matching, in which both an entity’s boundary and category had to be correct, from token-level precision, recall and F1 calculated under the BIO scheme. The headline 96.52% figure is an exact-match measure across the entity categories, not a token-level F1 score. Performance differed by category and length.

Equipment and time entities achieved reported F1 scores of 99.2% and 98.6%, respectively, while operational-data entities reached 90.6%. The authors associated the lower operational-data result with numerous numerical variations and combinations of values and units. Entities of one or two words had an F1 score of 98.8%, compared with 88.4% for entities containing seven or more words.

In a manual analysis of 300 test samples containing operational-data or person-and-organization entities, 52.3% of operational-data errors were boundary errors, particularly around combined values and units. Among person-and-organization errors, 44.7% were category-confusion cases. The full model recorded a macro-average F1 of 94.62% in the ablation comparison, illustrating why the paper reports both exact-match accuracy and token-level measures.

The proposed system was compared with CNN-LSTM, thesaurus-based, Transformer-CRF, BiLSTM-CRF, CRF-only, BERT-only and RoBERTa-only baselines trained and tested on the same data and under the same experimental environment. In a separate architectural comparison, the plain BiGRU-CRF model had 8.7 million parameters, a training speed of 125 samples per second and an inference speed of 208 samples per second. It achieved 98.2% on medium-length sequences and 98.9% on structured data.

The Transformer-CRF used 15.3 million parameters and was reported as approximately 45% slower in training and 54% slower in inference. The BiLSTM-CRF reached an 88.7% F1 score on ultra-long sequences, whereas the CNN-BiGRU-CRF variant approached the BiGRU-CRF in several results without a clear advantage. These measurements apply to the study’s configurations and do not establish that the same ranking or speed will occur in other systems.

The reported timing results depended on an Intel Xeon Gold 6248R processor, 128 GB of memory, a batch size of 16, sequences shorter than 162 characters and GPU acceleration. Preprocessing averaged 5.2 milliseconds, while optimized tag decoding was kept within 2.1 milliseconds, and dynamic batching was used during inference. The authors presented cloud batch processing, scheduling-center servers and compressed edge models as possible deployment configurations, but described them as requiring further engineering validation.

They also identified explainability, cross-domain transfer and systematic evaluation after adding new entity categories as open questions. Overall, the study shows strong experimental performance for structured Chinese power-grid documents, while its deployment estimates remain conditional on hardware, sequence length, data characteristics and optimization rather than demonstrated outcomes in live power-system operations.

Subject of Research: Experimental evaluation of a BiGRU-CRF and NLP pipeline for entity recognition in Chinese power-grid text.

Article Title: Power data entity recognition method based on BiGRU-CRF and NLP

Article References: Gao, M., & Wang, L. (2026). Power data entity recognition method based on BiGRU-CRF and NLP. Discover Artificial Intelligence, 6(1), Article 1085. https://doi.org/10.1007/s44163-026-01789-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01789-w

Keywords: BiGRU, conditional random fields, power-grid text, named entity recognition, natural language processing, word segmentation, deep learning, Chinese language, power-system data, Power, data, entity